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

macOS TCC blocks mDNS multicast from SSH sessions, preventing SSH-spawned exo processes from discovering peers. This adds a --bootstrap-peer IP:PORT flag that directly dials a known peer, bypassing mDNS. The flag is additive — mDNS still runs alongside. Changes across Rust networking, PyO3 bindings, and Python CLI: - discovery.rs: store bootstrap addresses, dial on startup and retry - swarm.rs: thread bootstrap_peers through to Behaviour - networking.rs: accept optional bootstrap_peers in PyO3 constructor - router.py: pass bootstrap_peers to NetworkingHandle - main.py: parse --bootstrap-peer CLI arg into multiaddr format Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
fix: cancel active tasks on meta-instance cascade delete
2026-02-16 10:24:00 -08:00 · 2026-02-16 09:51:55 -08:00 · 2026-02-16 05:38:47 -08:00 · 2026-02-16 05:24:49 -08:00 · 2026-02-16 05:11:29 -08:00 · 2026-02-15 16:13:38 -08:00
167 changed files with 7083 additions and 9960 deletions
@@ -200,7 +200,7 @@ class Module(dict):
    ) -> mx.MX_ARRAY_TREE:  # -> dict[Any, Any | dict[Any, Any | dict[Any, Any] | list[Any]] | dict[Any, Any] | list[Any]]:
        """Return the submodules that do not contain other modules."""

-    def update(self, parameters: dict[str, Any], strict: bool = ...) -> Module:
+    def update(self, parameters: dict, strict: bool = ...) -> Module:
        """Replace the parameters of this Module with the provided ones in the
        dict of dicts and lists.

@@ -7,10 +7,7 @@ from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 from mlx.core import MX_ARRAY_TREE

 def tree_map(
-    fn: Callable[..., Any],
-    tree: Any,
-    *rest: Any,
-    is_leaf: Callable[..., bool] | None = ...,
+    fn: Callable, tree: Any, *rest: Any, is_leaf: Optional[Callable] = ...
 ) -> Any:
    """Applies ``fn`` to the leaves of the Python tree ``tree`` and
    returns a new collection with the results.
@@ -47,11 +44,11 @@ def tree_map(
    """

 def tree_map_with_path(
-    fn: Callable[..., Any],
+    fn: Callable,
    tree: Any,
    *rest: Any,
-    is_leaf: Callable[..., bool] | None = ...,
-    path: str | None = ...,
+    is_leaf: Optional[Callable] = ...,
+    path: Optional[Any] = ...,
 ) -> Any:
    """Applies ``fn`` to the path and leaves of the Python tree ``tree`` and
    returns a new collection with the results.
@@ -83,9 +80,9 @@ def tree_map_with_path(
 def tree_flatten(
    tree: Any,
    prefix: str = ...,
-    is_leaf: Callable[..., bool] | None = ...,
-    destination: list[tuple[str, Any]] | dict[str, Any] | None = ...,
-) -> list[tuple[str, Any]] | dict[str, Any]:
+    is_leaf: Optional[Callable] = ...,
+    destination: Optional[Union[List[Tuple[str, Any]], Dict[str, Any]]] = ...,
+) -> Union[List[Tuple[str, Any]], Dict[str, Any]]:
    """Flattens a Python tree to a list of key, value tuples.

    The keys are using the dot notation to define trees of arbitrary depth and
@@ -121,7 +118,7 @@ def tree_flatten(
            the Python tree.
    """

-def tree_unflatten(tree: list[tuple[str, Any]] | dict[str, Any]) -> Any:
+def tree_unflatten(tree: Union[List[Tuple[str, Any]], Dict[str, Any]]) -> Any:
    """Recreate a Python tree from its flat representation.

    .. code-block:: python
@@ -1,46 +0,0 @@
-"""Type stubs for mlx_lm.models.glm_moe_dsa"""
-
-from dataclasses import dataclass
-from typing import Any, Dict, Optional
-
-from .base import BaseModelArgs
-from .deepseek_v32 import Model as DSV32Model
-
-@dataclass
-class ModelArgs(BaseModelArgs):
-    model_type: str
-    vocab_size: int
-    hidden_size: int
-    index_head_dim: int
-    index_n_heads: int
-    index_topk: int
-    intermediate_size: int
-    moe_intermediate_size: int
-    num_hidden_layers: int
-    num_attention_heads: int
-    num_key_value_heads: int
-    n_shared_experts: Optional[int]
-    n_routed_experts: Optional[int]
-    routed_scaling_factor: float
-    kv_lora_rank: int
-    q_lora_rank: int
-    qk_rope_head_dim: int
-    v_head_dim: int
-    qk_nope_head_dim: int
-    topk_method: str
-    scoring_func: str
-    norm_topk_prob: bool
-    n_group: int
-    topk_group: int
-    num_experts_per_tok: int
-    moe_layer_freq: int
-    first_k_dense_replace: int
-    max_position_embeddings: int
-    rms_norm_eps: float
-    rope_parameters: Dict[str, Any]
-    attention_bias: bool
-    rope_scaling: Dict[str, Any] | None
-    rope_theta: float | None
-
-class Model(DSV32Model):
-    def __init__(self, config: ModelArgs) -> None: ...
@@ -141,6 +141,12 @@ version = "0.3.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "76a2e8124351fda1ef8aaaa3bbd7ebbcb486bbcd4225aca0aa0d84bb2db8fecb"

+[[package]]
+name = "arrayvec"
+version = "0.7.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7c02d123df017efcdfbd739ef81735b36c5ba83ec3c59c80a9d7ecc718f92e50"
+
 [[package]]
 name = "asn1-rs"
 version = "0.7.1"
@@ -298,6 +304,19 @@ version = "1.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "55248b47b0caf0546f7988906588779981c43bb1bc9d0c44087278f80cdb44ba"

+[[package]]
+name = "bigdecimal"
+version = "0.4.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "560f42649de9fa436b73517378a147ec21f6c997a546581df4b4b31677828934"
+dependencies = [
+ "autocfg",
+ "libm",
+ "num-bigint",
+ "num-integer",
+ "num-traits",
+]
+
 [[package]]
 name = "bimap"
 version = "0.6.3"
@@ -497,6 +516,15 @@ version = "0.4.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2f421161cb492475f1661ddc9815a745a1c894592070661180fdec3d4872e9c3"

+[[package]]
+name = "convert_case"
+version = "0.10.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "633458d4ef8c78b72454de2d54fd6ab2e60f9e02be22f3c6104cdc8a4e0fceb9"
+dependencies = [
+ "unicode-segmentation",
+]
+
 [[package]]
 name = "core-foundation"
 version = "0.9.4"
@@ -718,6 +746,29 @@ dependencies = [
 "powerfmt",
 ]

+[[package]]
+name = "derive_more"
+version = "2.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "10b768e943bed7bf2cab53df09f4bc34bfd217cdb57d971e769874c9a6710618"
+dependencies = [
+ "derive_more-impl",
+]
+
+[[package]]
+name = "derive_more-impl"
+version = "2.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6d286bfdaf75e988b4a78e013ecd79c581e06399ab53fbacd2d916c2f904f30b"
+dependencies = [
+ "convert_case",
+ "proc-macro2",
+ "quote",
+ "rustc_version",
+ "syn 2.0.111",
+ "unicode-xid",
+]
+
 [[package]]
 name = "digest"
 version = "0.10.7"
@@ -888,17 +939,22 @@ name = "exo_pyo3_bindings"
 version = "0.0.1"
 dependencies = [
 "delegate",
+ "derive_more",
 "env_logger",
 "extend",
- "futures-lite",
+ "futures",
+ "impl-trait-for-tuples",
 "libp2p",
 "log",
 "networking",
+ "once_cell",
 "pin-project",
 "pyo3",
 "pyo3-async-runtimes",
 "pyo3-log",
 "pyo3-stub-gen",
+ "thiserror 2.0.17",
+ "thread_local",
 "tokio",
 "util",
 ]
@@ -914,12 +970,6 @@ dependencies = [
 "syn 2.0.111",
 ]

-[[package]]
-name = "fastrand"
-version = "2.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "37909eebbb50d72f9059c3b6d82c0463f2ff062c9e95845c43a6c9c0355411be"
-
 [[package]]
 name = "ff"
 version = "0.13.1"
@@ -1028,10 +1078,7 @@ version = "2.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f78e10609fe0e0b3f4157ffab1876319b5b0db102a2c60dc4626306dc46b44ad"
 dependencies = [
- "fastrand",
 "futures-core",
- "futures-io",
- "parking",
 "pin-project-lite",
 ]

@@ -1593,6 +1640,17 @@ dependencies = [
 "xmltree",
 ]

+[[package]]
+name = "impl-trait-for-tuples"
+version = "0.2.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a0eb5a3343abf848c0984fe4604b2b105da9539376e24fc0a3b0007411ae4fd9"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.111",
+]
+
 [[package]]
 name = "indexmap"
 version = "2.12.1"
@@ -1771,6 +1829,12 @@ version = "0.2.178"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "37c93d8daa9d8a012fd8ab92f088405fb202ea0b6ab73ee2482ae66af4f42091"

+[[package]]
+name = "libm"
+version = "0.2.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f9fbbcab51052fe104eb5e5d351cf728d30a5be1fe14d9be8a3b097481fb97de"
+
 [[package]]
 name = "libp2p"
 version = "0.56.0"
@@ -2760,13 +2824,16 @@ name = "networking"
 version = "0.0.1"
 dependencies = [
 "delegate",
+ "derive_more",
 "either",
 "extend",
- "futures-lite",
+ "futures",
 "futures-timer",
+ "impl-trait-for-tuples",
 "keccak-const",
 "libp2p",
 "log",
+ "thiserror 2.0.17",
 "tokio",
 "tracing-subscriber",
 "util",
@@ -2851,6 +2918,17 @@ dependencies = [
 "num-traits",
 ]

+[[package]]
+name = "num-rational"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f83d14da390562dca69fc84082e73e548e1ad308d24accdedd2720017cb37824"
+dependencies = [
+ "num-bigint",
+ "num-integer",
+ "num-traits",
+]
+
 [[package]]
 name = "num-traits"
 version = "0.2.19"
@@ -3201,14 +3279,28 @@ version = "0.27.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ab53c047fcd1a1d2a8820fe84f05d6be69e9526be40cb03b73f86b6b03e6d87d"
 dependencies = [
+ "bigdecimal",
+ "either",
+ "hashbrown 0.16.1",
+ "indexmap",
 "indoc",
+ "inventory",
 "libc",
+ "lock_api",
 "memoffset",
+ "num-bigint",
+ "num-complex",
+ "num-rational",
+ "num-traits",
 "once_cell",
+ "ordered-float",
+ "parking_lot",
 "portable-atomic",
 "pyo3-build-config",
 "pyo3-ffi",
 "pyo3-macros",
+ "rust_decimal",
+ "smallvec",
 "unindent",
 ]

@@ -3649,6 +3741,16 @@ dependencies = [
 "tokio",
 ]

+[[package]]
+name = "rust_decimal"
+version = "1.39.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "35affe401787a9bd846712274d97654355d21b2a2c092a3139aabe31e9022282"
+dependencies = [
+ "arrayvec",
+ "num-traits",
+]
+
 [[package]]
 name = "rustc-hash"
 version = "1.1.0"
@@ -4513,12 +4615,24 @@ version = "1.0.22"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9312f7c4f6ff9069b165498234ce8be658059c6728633667c526e27dc2cf1df5"

+[[package]]
+name = "unicode-segmentation"
+version = "1.12.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f6ccf251212114b54433ec949fd6a7841275f9ada20dddd2f29e9ceea4501493"
+
 [[package]]
 name = "unicode-width"
 version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b4ac048d71ede7ee76d585517add45da530660ef4390e49b098733c6e897f254"

+[[package]]
+name = "unicode-xid"
+version = "0.2.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ebc1c04c71510c7f702b52b7c350734c9ff1295c464a03335b00bb84fc54f853"
+
 [[package]]
 name = "unicode_names2"
 version = "1.3.0"
@@ -26,20 +26,49 @@ opt-level = 3
 networking = { path = "rust/networking" }
 util = { path = "rust/util" }

+# Proc-macro authoring tools
+syn = "2.0"
+quote = "1.0"
+proc-macro2 = "1.0"
+darling = "0.20"
+
 # Macro dependecies
 extend = "1.2"
 delegate = "0.13"
+impl-trait-for-tuples = "0.2"
+clap = "4.5"
+derive_more = { version = "2.0.1", features = ["display"] }
+pin-project = "1"

 # Utility dependencies
+itertools = "0.14"
+thiserror = "2"
+internment = "0.8"
+recursion = "0.5"
+regex = "1.11"
+once_cell = "1.21"
+thread_local = "1.1"
+bon = "3.4"
+generativity = "1.1"
+anyhow = "1.0"
 keccak-const = "0.2"

+# Functional generics/lenses frameworks
+frunk_core = "0.4"
+frunk = "0.4"
+frunk_utils = "0.2"
+frunk-enum-core = "0.3"
+
 # Async dependencies
 tokio = "1.46"
-futures-lite = "2.6.1"
+futures = "0.3"
+futures-util = "0.3"
 futures-timer = "3.0"

 # Data structures
 either = "1.15"
+ordered-float = "5.0"
+ahash = "0.8"

 # Tracing/logging
 log = "0.4"
@@ -72,30 +72,16 @@ There are two ways to run exo:

 ### Run from Source (macOS)

-If you have [Nix](https://nixos.org/) installed, you can skip most of the steps below and run exo directly:
-
-```bash
-nix run .#exo
-```
-
-**Note:** To accept the Cachix binary cache (and avoid the Xcode Metal ToolChain), add to `/etc/nix/nix.conf`:
-```
-trusted-users = root    (or your username)
-experimental-features = nix-command flakes
-```
-Then restart the Nix daemon: `sudo launchctl kickstart -k system/org.nixos.nix-daemon`
-
 **Prerequisites:**
- [Xcode](https://developer.apple.com/xcode/) (provides the Metal ToolChain required for MLX compilation)
 - [brew](https://github.com/Homebrew/brew) (for simple package management on macOS)
-
+  
  ```bash
  /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
  ```
 - [uv](https://github.com/astral-sh/uv) (for Python dependency management)
 - [macmon](https://github.com/vladkens/macmon) (for hardware monitoring on Apple Silicon)
 - [node](https://github.com/nodejs/node) (for building the dashboard)
-
+  
  ```bash
  brew install uv macmon node
  ```
@@ -126,37 +126,11 @@ final class ExoProcessController: ObservableObject {
            return
        }
        process.terminationHandler = nil
-        status = .stopped
-
-        guard process.isRunning else {
-            self.process = nil
-            return
+        if process.isRunning {
+            process.terminate()
        }
-
-        let proc = process
        self.process = nil
-
-        Task.detached {
-            proc.interrupt()
-
-            for _ in 0..<50 {
-                if !proc.isRunning { return }
-                try? await Task.sleep(nanoseconds: 100_000_000)
-            }
-
-            if proc.isRunning {
-                proc.terminate()
-            }
-
-            for _ in 0..<30 {
-                if !proc.isRunning { return }
-                try? await Task.sleep(nanoseconds: 100_000_000)
-            }
-
-            if proc.isRunning {
-                kill(proc.processIdentifier, SIGKILL)
-            }
-        }
+        status = .stopped
    }

    func restart() {
@@ -1,7 +0,0 @@
-# Canary benchmark manifest
-#
-# Lists the suite files to include. Each file defines benchmarks
-# with shared constraints, topology, and default args.
-include = [
-    "single-m3-ultra.toml",
-]
@@ -1,48 +1,29 @@
-# type: ignore
 #!/usr/bin/env python3
-"""Tool-calling eval for exo's OpenAI-compatible API.
-
-Tests whether models correctly:
- Trigger tool calls when appropriate
- Return valid JSON arguments matching function schemas
- Handle multi-turn tool use (call -> result -> final answer)
- Avoid calling tools when unnecessary
-
-Start exo with a model first, then run:
-    uv run python tool_call_eval.py --model <model-id>
-    uv run python tool_call_eval.py --model <model-id> --host 10.0.0.5 --port 52415
-    uv run python tool_call_eval.py --model <model-id> --repeat 3
-    uv run python tool_call_eval.py --model <model-id> --scenarios weather_simple calculator_multi_turn
-"""
-
+# pyright: reportAny=false, reportUnknownMemberType=false, reportUnknownVariableType=false, reportUnknownArgumentType=false
 from __future__ import annotations

 import argparse
 import contextlib
+import http.client
 import itertools
 import json
+import os
 import sys
 import time
 from collections.abc import Callable
 from pathlib import Path
 from statistics import mean
 from typing import Any
+from urllib.parse import urlencode

-from harness import (
-    ExoClient,
-    ExoHttpError,
-    add_common_instance_args,
-    instance_id_from_instance,
-    nodes_used_in_instance,
-    resolve_model_short_id,
-    run_planning_phase,
-    settle_and_fetch_placements,
-    wait_for_instance_gone,
-    wait_for_instance_ready,
-)
 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
@@ -122,6 +103,154 @@ def load_tokenizer_for_bench(model_id: str) -> Any:
    return AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)


+class ExoHttpError(RuntimeError):
+    def __init__(self, status: int, reason: str, body_preview: str):
+        super().__init__(f"HTTP {status} {reason}: {body_preview}")
+        self.status = status
+
+
+class ExoClient:
+    def __init__(self, host: str, port: int, timeout_s: float = 7200.0):
+        self.host = host
+        self.port = port
+        self.timeout_s = timeout_s
+
+    def request_json(
+        self,
+        method: str,
+        path: str,
+        params: dict[str, Any] | None = None,
+        body: dict[str, Any] | None = None,
+        headers: dict[str, str] | None = None,
+    ) -> Any:
+        if not path.startswith("/"):
+            path = "/" + path
+        if params:
+            path = path + "?" + urlencode(params)
+
+        conn = http.client.HTTPConnection(self.host, self.port, timeout=self.timeout_s)
+        try:
+            payload: bytes | None = None
+            hdrs: dict[str, str] = {"Accept": "application/json"}
+
+            if body is not None:
+                payload = json.dumps(body).encode("utf-8")
+                hdrs["Content-Type"] = "application/json"
+            if headers:
+                hdrs.update(headers)
+
+            conn.request(method.upper(), path, body=payload, headers=hdrs)
+            resp = conn.getresponse()
+            raw = resp.read()
+            text = raw.decode("utf-8", errors="replace") if raw else ""
+
+            if resp.status >= 400:
+                raise ExoHttpError(resp.status, resp.reason, text[:300])
+
+            if not text:
+                return None
+            return json.loads(text)
+        finally:
+            conn.close()
+
+    def post_bench_chat_completions(self, payload: dict[str, Any]) -> dict[str, Any]:
+        return self.request_json("POST", "/bench/chat/completions", body=payload)
+
+
+def unwrap_instance(instance: dict[str, Any]) -> dict[str, Any]:
+    if len(instance) != 1:
+        raise KeyError(f"Expected 1 key, got keys={list(instance.keys())}")
+
+    tag = next(iter(instance))
+    inner = instance[tag]
+    if not isinstance(inner, dict):
+        raise TypeError(f"payload for {tag} must be dict, got {type(inner)}")
+    return inner
+
+
+def instance_id_from_instance(instance: dict[str, Any]) -> str:
+    inner = unwrap_instance(instance)
+    return str(inner["instanceId"])
+
+
+def nodes_used_in_instance(instance: dict[str, Any]) -> int:
+    inner = unwrap_instance(instance)
+    return len(inner["shardAssignments"]["nodeToRunner"])
+
+
+def runner_ids_from_instance(instance: dict[str, Any]) -> list[str]:
+    inner = unwrap_instance(instance)
+    runner_to_shard = inner["shardAssignments"]["runnerToShard"]
+    return list(runner_to_shard.keys())
+
+
+def runner_ready(runner: dict[str, Any]) -> bool:
+    return "RunnerReady" in runner
+
+
+def runner_failed(runner: dict[str, Any]) -> bool:
+    return "RunnerFailed" in runner
+
+
+def get_runner_failed_message(runner: dict[str, Any]) -> str | None:
+    if "RunnerFailed" in runner:
+        return runner["RunnerFailed"].get("errorMessage")
+    return None
+
+
+def wait_for_instance_ready(
+    client: ExoClient, instance_id: str, timeout: float = 24000.0
+) -> None:
+    start_time = time.time()
+    instance_existed = False
+    while time.time() - start_time < timeout:
+        state = client.request_json("GET", "/state")
+        instances = state.get("instances", {})
+
+        if instance_id not in instances:
+            if instance_existed:
+                # Instance was deleted after being created - likely due to runner failure
+                raise RuntimeError(
+                    f"Instance {instance_id} was deleted (runner may have failed)"
+                )
+            time.sleep(0.1)
+            continue
+
+        instance_existed = True
+        instance = instances[instance_id]
+        runner_ids = runner_ids_from_instance(instance)
+        runners = state.get("runners", {})
+
+        # Check for failed runners first
+        for rid in runner_ids:
+            runner = runners.get(rid, {})
+            if runner_failed(runner):
+                error_msg = get_runner_failed_message(runner) or "Unknown error"
+                raise RuntimeError(f"Runner {rid} failed: {error_msg}")
+
+        if all(runner_ready(runners.get(rid, {})) for rid in runner_ids):
+            return
+
+        time.sleep(0.1)
+
+    raise TimeoutError(f"Instance {instance_id} did not become ready within {timeout=}")
+
+
+def wait_for_instance_gone(
+    client: ExoClient, instance_id: str, timeout: float = 3.0
+) -> None:
+    start_time = time.time()
+    while time.time() - start_time < timeout:
+        try:
+            client.request_json("GET", f"/instance/{instance_id}")
+            time.sleep(0.4)
+        except ExoHttpError as e:
+            if e.status == 404:
+                return
+
+    raise TimeoutError(f"Instance {instance_id} did not get deleted within {timeout=}")
+
+
 def format_peak_memory(b: float) -> str:
    for unit in ["B", "KB", "MB", "GB", "TB"]:
        if b < 1024.0:
@@ -140,6 +269,184 @@ def parse_int_list(values: list[str]) -> list[int]:
    return items


+def resolve_model_short_id(client: ExoClient, model_arg: str) -> tuple[str, str]:
+    models = client.request_json("GET", "/models") or {}
+    data = models.get("data") or []
+
+    for m in data:
+        if m.get("name").lower() == model_arg.lower():
+            short_id = str(m["name"])
+            full_id = str(m.get("hugging_face_id") or m["name"])
+            return short_id, full_id
+
+    for m in data:
+        if m.get("hugging_face_id") == model_arg:
+            short_id = str(m["name"])
+            full_id = str(m["hugging_face_id"])
+            return short_id, full_id
+
+    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":
+        return ("ring" in s) or ("jaccl" in s)
+    return wanted in s
+
+
+def sharding_filter(sharding: str, wanted: str) -> bool:
+    s = (sharding or "").lower()
+    if wanted == "both":
+        return ("pipeline" in s) or ("tensor" in s)
+    return wanted in s
+
+
 def run_one_completion(
    client: ExoClient, model_id: str, pp_hint: int, tg: int, prompt_sizer: PromptSizer
 ) -> tuple[dict[str, Any], int]:
@@ -231,12 +538,76 @@ 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",
        description="Benchmark exo model throughput across placement previews.",
    )
-    add_common_instance_args(ap)
+    ap.add_argument("--host", default=os.environ.get("EXO_HOST", "localhost"))
+    ap.add_argument(
+        "--port", type=int, default=int(os.environ.get("EXO_PORT", "52415"))
+    )
+    ap.add_argument("--model", required=True, help="Model short id or huggingface id")
    ap.add_argument(
        "--pp",
        nargs="+",
@@ -249,6 +620,34 @@ def main() -> int:
        required=True,
        help="Generation lengths (ints). Accepts commas.",
    )
+    ap.add_argument(
+        "--max-nodes",
+        type=int,
+        default=4,
+        help="Only consider placements using <= this many nodes.",
+    )
+    ap.add_argument(
+        "--min-nodes",
+        type=int,
+        default=1,
+        help="Only consider placements using >= this many nodes.",
+    )
+    ap.add_argument(
+        "--instance-meta", choices=["ring", "jaccl", "both"], default="both"
+    )
+    ap.add_argument(
+        "--sharding", choices=["pipeline", "tensor", "both"], default="both"
+    )
+    ap.add_argument(
+        "--skip-pipeline-jaccl",
+        action="store_true",
+        help="Skip pipeline+jaccl placements, as it's often pointless.",
+    )
+    ap.add_argument(
+        "--skip-tensor-ring",
+        action="store_true",
+        help="Skip tensor+ring placements, as it's so slow.",
+    )
    ap.add_argument(
        "--repeat", type=int, default=1, help="Repetitions per (pp,tg) pair."
    )
@@ -258,6 +657,9 @@ def main() -> int:
        default=0,
        help="Warmup runs per placement (uses first pp/tg).",
    )
+    ap.add_argument(
+        "--timeout", type=float, default=7200.0, help="HTTP timeout (seconds)."
+    )
    ap.add_argument(
        "--json-out",
        default="bench/results.json",
@@ -272,6 +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)
@@ -306,10 +719,24 @@ def main() -> int:
        logger.error("[exo-bench] tokenizer usable but prompt sizing failed")
        raise

-    selected = settle_and_fetch_placements(
-        client, full_model_id, args, settle_timeout=args.settle_timeout
+    settle_deadline = (
+        time.monotonic() + args.settle_timeout if args.settle_timeout > 0 else None
    )

+    selected = fetch_and_filter_placements(client, full_model_id, args)
+
+    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)..."
+            )
+            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.")
        return 1
@@ -333,12 +760,8 @@ def main() -> int:
    if args.dry_run:
        return 0

-    settle_deadline = (
-        time.monotonic() + args.settle_timeout if args.settle_timeout > 0 else None
-    )
-
    logger.info("Planning phase: checking downloads...")
-    download_duration_s = run_planning_phase(
+    run_planning_phase(
        client,
        full_model_id,
        selected[0],
@@ -346,10 +769,6 @@ def main() -> int:
        args.timeout,
        settle_deadline,
    )
-    if download_duration_s is not None:
-        logger.info(f"Download: {download_duration_s:.1f}s (freshly downloaded)")
-    else:
-        logger.info("Download: model already cached")

    all_rows: list[dict[str, Any]] = []

@@ -413,11 +832,6 @@ def main() -> int:
                            "pp_tokens": actual_pp_tokens,
                            "tg": tg,
                            "repeat_index": r,
-                            **(
-                                {"download_duration_s": download_duration_s}
-                                if download_duration_s is not None
-                                else {}
-                            ),
                        }
                    )
                    runs.append(row)
@@ -1,489 +0,0 @@
-# type: ignore
-from __future__ import annotations
-
-import argparse
-import http.client
-import json
-import os
-import time
-from typing import Any
-from urllib.parse import urlencode
-
-from loguru import logger
-
-_SETTLE_INITIAL_BACKOFF_S = 1.0
-_SETTLE_MAX_BACKOFF_S = 60.0
-_SETTLE_BACKOFF_MULTIPLIER = 2.0
-
-
-class ExoHttpError(RuntimeError):
-    def __init__(self, status: int, reason: str, body_preview: str):
-        super().__init__(f"HTTP {status} {reason}: {body_preview}")
-        self.status = status
-
-
-class ExoClient:
-    def __init__(self, host: str, port: int, timeout_s: float = 7200.0):
-        self.host = host
-        self.port = port
-        self.timeout_s = timeout_s
-
-    def request_json(
-        self,
-        method: str,
-        path: str,
-        params: dict[str, Any] | None = None,
-        body: dict[str, Any] | None = None,
-        headers: dict[str, str] | None = None,
-    ) -> Any:
-        if not path.startswith("/"):
-            path = "/" + path
-        if params:
-            path = path + "?" + urlencode(params)
-
-        conn = http.client.HTTPConnection(self.host, self.port, timeout=self.timeout_s)
-        try:
-            payload: bytes | None = None
-            hdrs: dict[str, str] = {"Accept": "application/json"}
-
-            if body is not None:
-                payload = json.dumps(body).encode("utf-8")
-                hdrs["Content-Type"] = "application/json"
-            if headers:
-                hdrs.update(headers)
-
-            conn.request(method.upper(), path, body=payload, headers=hdrs)
-            resp = conn.getresponse()
-            raw = resp.read()
-            text = raw.decode("utf-8", errors="replace") if raw else ""
-
-            if resp.status >= 400:
-                raise ExoHttpError(resp.status, resp.reason, text[:300])
-
-            if not text:
-                return None
-            return json.loads(text)
-        finally:
-            conn.close()
-
-    def post_bench_chat_completions(self, payload: dict[str, Any]) -> dict[str, Any]:
-        return self.request_json("POST", "/bench/chat/completions", body=payload)
-
-
-def unwrap_instance(instance: dict[str, Any]) -> dict[str, Any]:
-    if len(instance) != 1:
-        raise KeyError(f"Expected 1 key, got keys={list(instance.keys())}")
-
-    tag = next(iter(instance))
-    inner = instance[tag]
-    if not isinstance(inner, dict):
-        raise TypeError(f"payload for {tag} must be dict, got {type(inner)}")
-    return inner
-
-
-def instance_id_from_instance(instance: dict[str, Any]) -> str:
-    inner = unwrap_instance(instance)
-    return str(inner["instanceId"])
-
-
-def nodes_used_in_instance(instance: dict[str, Any]) -> int:
-    inner = unwrap_instance(instance)
-    return len(inner["shardAssignments"]["nodeToRunner"])
-
-
-def runner_ids_from_instance(instance: dict[str, Any]) -> list[str]:
-    inner = unwrap_instance(instance)
-    runner_to_shard = inner["shardAssignments"]["runnerToShard"]
-    return list(runner_to_shard.keys())
-
-
-def runner_ready(runner: dict[str, Any]) -> bool:
-    return "RunnerReady" in runner
-
-
-def runner_failed(runner: dict[str, Any]) -> bool:
-    return "RunnerFailed" in runner
-
-
-def get_runner_failed_message(runner: dict[str, Any]) -> str | None:
-    if "RunnerFailed" in runner:
-        return runner["RunnerFailed"].get("errorMessage")
-    return None
-
-
-def wait_for_instance_ready(
-    client: ExoClient, instance_id: str, timeout: float = 24000.0
-) -> None:
-    start_time = time.time()
-    instance_existed = False
-    while time.time() - start_time < timeout:
-        state = client.request_json("GET", "/state")
-        instances = state.get("instances", {})
-
-        if instance_id not in instances:
-            if instance_existed:
-                # Instance was deleted after being created - likely due to runner failure
-                raise RuntimeError(
-                    f"Instance {instance_id} was deleted (runner may have failed)"
-                )
-            time.sleep(0.1)
-            continue
-
-        instance_existed = True
-        instance = instances[instance_id]
-        runner_ids = runner_ids_from_instance(instance)
-        runners = state.get("runners", {})
-
-        # Check for failed runners first
-        for rid in runner_ids:
-            runner = runners.get(rid, {})
-            if runner_failed(runner):
-                error_msg = get_runner_failed_message(runner) or "Unknown error"
-                raise RuntimeError(f"Runner {rid} failed: {error_msg}")
-
-        if all(runner_ready(runners.get(rid, {})) for rid in runner_ids):
-            return
-
-        time.sleep(0.1)
-
-    raise TimeoutError(f"Instance {instance_id} did not become ready within {timeout=}")
-
-
-def wait_for_instance_gone(
-    client: ExoClient, instance_id: str, timeout: float = 3.0
-) -> None:
-    start_time = time.time()
-    while time.time() - start_time < timeout:
-        try:
-            client.request_json("GET", f"/instance/{instance_id}")
-            time.sleep(0.4)
-        except ExoHttpError as e:
-            if e.status == 404:
-                return
-            raise
-
-    raise TimeoutError(f"Instance {instance_id} did not get deleted within {timeout=}")
-
-
-def resolve_model_short_id(client: ExoClient, model_arg: str) -> tuple[str, str]:
-    models = client.request_json("GET", "/models") or {}
-    data = models.get("data") or []
-
-    for m in data:
-        if (m.get("name") or "").lower() == model_arg.lower():
-            short_id = str(m["name"])
-            full_id = str(m.get("hugging_face_id") or m["name"])
-            return short_id, full_id
-
-    for m in data:
-        if m.get("hugging_face_id") == model_arg:
-            short_id = str(m["name"])
-            full_id = str(m["hugging_face_id"])
-            return short_id, full_id
-
-    raise ValueError(f"Model not found in /models: {model_arg}")
-
-
-def placement_filter(instance_meta: str, wanted: str) -> bool:
-    s = (instance_meta or "").lower()
-    if wanted == "both":
-        return ("ring" in s) or ("jaccl" in s)
-    return wanted in s
-
-
-def sharding_filter(sharding: str, wanted: str) -> bool:
-    s = (sharding or "").lower()
-    if wanted == "both":
-        return ("pipeline" in s) or ("tensor" in s)
-    return wanted in s
-
-
-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 settle_and_fetch_placements(
-    client: ExoClient,
-    full_model_id: str,
-    args: argparse.Namespace,
-    settle_timeout: float = 0,
-) -> list[dict[str, Any]]:
-    selected = fetch_and_filter_placements(client, full_model_id, args)
-
-    if not selected and settle_timeout > 0:
-        backoff = _SETTLE_INITIAL_BACKOFF_S
-        deadline = time.monotonic() + settle_timeout
-        while not selected and time.monotonic() < deadline:
-            remaining = 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)..."
-            )
-            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)
-
-    return selected
-
-
-def run_planning_phase(
-    client: ExoClient,
-    full_model_id: str,
-    preview: dict[str, Any],
-    danger_delete: bool,
-    timeout: float,
-    settle_deadline: float | None,
-) -> float | None:
-    """Check disk space and ensure model is downloaded before benchmarking.
-
-    Returns the wall-clock download duration in seconds if a fresh download
-    was needed, or None if the model was already cached on all nodes.
-    """
-    # 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 None
-
-    # 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", {})
-
-    needs_download = False
-
-    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
-
-        needs_download = True
-
-        # 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 (skip read-only models from EXO_MODELS_PATH)
-        completed = [
-            (
-                unwrap_instance(p["DownloadCompleted"]["shardMetadata"])["modelCard"][
-                    "modelId"
-                ],
-                p["DownloadCompleted"]["totalBytes"]["inBytes"],
-            )
-            for p in node_downloads
-            if "DownloadCompleted" in p
-            and not p["DownloadCompleted"].get("readOnly", False)
-        ]
-        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)
-    download_t0 = time.perf_counter() if needs_download else None
-    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:
-            if download_t0 is not None:
-                return time.perf_counter() - download_t0
-            return None
-        time.sleep(1)
-
-    raise TimeoutError("Downloads did not complete in time")
-
-
-def add_common_instance_args(ap: argparse.ArgumentParser) -> None:
-    ap.add_argument("--host", default=os.environ.get("EXO_HOST", "localhost"))
-    ap.add_argument(
-        "--port", type=int, default=int(os.environ.get("EXO_PORT", "52415"))
-    )
-    ap.add_argument("--model", required=True, help="Model short id or huggingface id")
-    ap.add_argument(
-        "--max-nodes",
-        type=int,
-        default=4,
-        help="Only consider placements using <= this many nodes.",
-    )
-    ap.add_argument(
-        "--min-nodes",
-        type=int,
-        default=1,
-        help="Only consider placements using >= this many nodes.",
-    )
-    ap.add_argument(
-        "--instance-meta", choices=["ring", "jaccl", "both"], default="both"
-    )
-    ap.add_argument(
-        "--sharding", choices=["pipeline", "tensor", "both"], default="both"
-    )
-    ap.add_argument(
-        "--skip-pipeline-jaccl",
-        action="store_true",
-        help="Skip pipeline+jaccl placements, as it's often pointless.",
-    )
-    ap.add_argument(
-        "--skip-tensor-ring",
-        action="store_true",
-        help="Skip tensor+ring placements, as it's so slow.",
-    )
-    ap.add_argument(
-        "--timeout", type=float, default=7200.0, help="HTTP timeout (seconds)."
-    )
-    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.",
-    )
@@ -4,7 +4,6 @@ version = "0.1.0"
 description = "Benchmarking tool for exo distributed inference"
 requires-python = ">=3.13"
 dependencies = [
-    "httpx>=0.27.0",
    "loguru>=0.7.3",
    "transformers>=5.0.0",
    "huggingface-hub>=0.33.4",
@@ -1,306 +0,0 @@
-# Tool definitions — each becomes an OpenAI function tool.
-# All scenarios get all tools unless they specify a `tools` list.
-
-[tools.get_current_weather]
-description = "Get the current weather in a given location"
-required = ["location"]
-
-[tools.get_current_weather.properties.location]
-type = "string"
-description = "City and state, e.g. San Francisco, CA"
-
-[tools.get_current_weather.properties.unit]
-type = "string"
-enum = ["celsius", "fahrenheit"]
-description = "Temperature unit"
-
-[tools.calculate]
-description = "Evaluate a mathematical expression and return the numeric result"
-required = ["expression"]
-
-[tools.calculate.properties.expression]
-type = "string"
-description = "The math expression to evaluate, e.g. '2 + 3 * 4'"
-
-[tools.search_products]
-description = "Search for products in a catalog by query, category, and price"
-required = ["query"]
-
-[tools.search_products.properties.query]
-type = "string"
-description = "Search query string"
-
-[tools.search_products.properties.category]
-type = "string"
-enum = ["electronics", "clothing", "food", "books"]
-description = "Product category to filter by"
-
-[tools.search_products.properties.max_price]
-type = "number"
-description = "Maximum price in USD"
-
-[tools.create_todos]
-description = "Create a structured todo list"
-required = ["todos"]
-
-[tools.create_todos.properties.todos]
-type = "array"
-description = "List of todo items"
-
-[tools.create_todos.properties.todos.items]
-type = "object"
-required = ["content", "status", "priority"]
-
-[tools.create_todos.properties.todos.items.properties.content]
-type = "string"
-description = "The todo item text"
-
-[tools.create_todos.properties.todos.items.properties.status]
-type = "string"
-description = "Status: pending, in_progress, or completed"
-
-[tools.create_todos.properties.todos.items.properties.priority]
-type = "string"
-description = "Priority: low, normal, or high"
-
-# -- Should call a tool --
-
-[[scenarios]]
-name = "weather_simple"
-description = "Basic weather query -> get_current_weather"
-expect_tool_call = true
-expected_function = "get_current_weather"
-required_arg_keys = ["location"]
-
-[[scenarios.messages]]
-role = "user"
-content = "What's the weather like in Tokyo right now?"
-
-[[scenarios]]
-name = "calculator_simple"
-description = "Math question -> calculate"
-expect_tool_call = true
-expected_function = "calculate"
-required_arg_keys = ["expression"]
-
-[[scenarios.messages]]
-role = "user"
-content = "Use the calculator to compute 3847 * 926 + 17293"
-
-[[scenarios]]
-name = "search_with_filters"
-description = "Product search with category and price filter"
-expect_tool_call = true
-expected_function = "search_products"
-required_arg_keys = ["query"]
-
-[[scenarios.messages]]
-role = "user"
-content = "Find me electronics under $50"
-
-# -- Multi-turn: tool call then follow-up --
-
-[[scenarios]]
-name = "weather_multi_turn"
-description = "Weather query -> tool result -> natural language summary"
-expect_tool_call = true
-expected_function = "get_current_weather"
-required_arg_keys = ["location"]
-
-[scenarios.tool_result]
-temperature = "18C"
-condition = "partly cloudy"
-humidity = "65%"
-wind = "12 km/h NW"
-
-[[scenarios.messages]]
-role = "user"
-content = "What's the weather in Paris?"
-
-[[scenarios]]
-name = "calculator_multi_turn"
-description = "Math query -> tool result -> model reports the answer"
-expect_tool_call = true
-expected_function = "calculate"
-required_arg_keys = ["expression"]
-
-[scenarios.tool_result]
-result = 491682
-
-[[scenarios.messages]]
-role = "user"
-content = "Use the calculator to compute 1847 * 263 + 5921"
-
-[[scenarios]]
-name = "search_multi_turn"
-description = "Search query -> tool result -> model summarizes products"
-expect_tool_call = true
-expected_function = "search_products"
-required_arg_keys = ["query"]
-
-[[scenarios.tool_result.results]]
-name = "Hands-On Machine Learning"
-price = 45.99
-rating = 4.8
-
-[[scenarios.tool_result.results]]
-name = "Deep Learning with Python"
-price = 39.99
-rating = 4.6
-
-[[scenarios.messages]]
-role = "user"
-content = "Search for books about machine learning"
-
-# -- Sequential tool calls --
-
-[[scenarios]]
-name = "chained_tool_calls_same"
-description = "Thinking + weather(Tokyo) -> result -> model must call weather(London)"
-expect_tool_call = true
-expected_function = "get_current_weather"
-required_arg_keys = ["location"]
-
-[[scenarios.messages]]
-role = "user"
-content = "Compare the weather in Tokyo and London."
-
-[[scenarios.messages]]
-role = "assistant"
-content = "I'll check both cities. Let me start with Tokyo."
-
-[[scenarios.messages.tool_calls]]
-id = "call_1"
-name = "get_current_weather"
-arguments = { location = "Tokyo" }
-
-[[scenarios.messages]]
-role = "tool"
-tool_call_id = "call_1"
-content = '{"temperature": "25C", "condition": "sunny"}'
-
-[[scenarios]]
-name = "chained_tool_calls_different"
-description = "Thinking + weather(Berlin) -> result -> model must call calculator"
-expect_tool_call = true
-expected_function = "calculate"
-required_arg_keys = ["expression"]
-
-[[scenarios.messages]]
-role = "user"
-content = "What's the weather in Berlin, and also use the calculator to compute 4819 * 37 + 291."
-
-[[scenarios.messages]]
-role = "assistant"
-content = "I'll handle both. Let me check Berlin's weather first."
-
-[[scenarios.messages.tool_calls]]
-id = "call_2"
-name = "get_current_weather"
-arguments = { location = "Berlin" }
-
-[[scenarios.messages]]
-role = "tool"
-tool_call_id = "call_2"
-content = '{"temperature": "12C", "condition": "rainy"}'
-
-[[scenarios]]
-name = "chained_tool_calls_three"
-description = "Two prior thinking+tool calls -> results -> model must make a third"
-expect_tool_call = true
-expected_function = "get_current_weather"
-required_arg_keys = ["location"]
-
-[[scenarios.messages]]
-role = "user"
-content = "Compare weather in Tokyo, Paris, and London."
-
-[[scenarios.messages]]
-role = "assistant"
-content = "I'll check all three cities. Starting with Tokyo."
-
-[[scenarios.messages.tool_calls]]
-id = "call_3"
-name = "get_current_weather"
-arguments = { location = "Tokyo" }
-
-[[scenarios.messages]]
-role = "tool"
-tool_call_id = "call_3"
-content = '{"temperature": "25C", "condition": "sunny"}'
-
-[[scenarios.messages]]
-role = "assistant"
-content = "Got Tokyo. Now checking Paris."
-
-[[scenarios.messages.tool_calls]]
-id = "call_4"
-name = "get_current_weather"
-arguments = { location = "Paris" }
-
-[[scenarios.messages]]
-role = "tool"
-tool_call_id = "call_4"
-content = '{"temperature": "18C", "condition": "cloudy"}'
-
-# -- Nested object schema (regression for lossy chat template rendering) --
-
-[[scenarios]]
-name = "nested_schema_tool_call"
-description = "Tool call with nested object array schema -> create_todos"
-expect_tool_call = true
-expected_function = "create_todos"
-required_arg_keys = ["todos"]
-nested_array_key = "todos"
-required_item_keys = ["content", "status", "priority"]
-tools = ["create_todos"]
-
-[[scenarios.messages]]
-role = "user"
-content = "Create a todo list with 3 items to learn Python"
-
-# -- Tool name integrity (regression for harmony token leaking into name) --
-
-[tools.glob]
-description = "Search for files matching a glob pattern in the codebase"
-required = ["pattern"]
-
-[tools.glob.properties.pattern]
-type = "string"
-description = "The glob pattern to match files against, e.g. '**/*.py'"
-
-[tools.glob.properties.path]
-type = "string"
-description = "The directory to search in"
-
-[[scenarios]]
-name = "tool_name_integrity"
-description = "Tool name must not contain harmony tokens like <|channel|>"
-expect_tool_call = true
-expected_function = "glob"
-required_arg_keys = ["pattern"]
-tools = ["glob"]
-
-[[scenarios.messages]]
-role = "user"
-content = "Find all Python files in the src directory"
-
-# -- Should NOT call a tool --
-
-[[scenarios]]
-name = "no_tool_joke"
-description = "Joke request should NOT trigger any tool"
-expect_tool_call = false
-
-[[scenarios.messages]]
-role = "user"
-content = "Tell me a funny joke about cats."
-
-[[scenarios]]
-name = "no_tool_factual"
-description = "Factual question answerable from training data"
-expect_tool_call = false
-
-[[scenarios.messages]]
-role = "user"
-content = "What is the capital of Japan?"
@@ -1,189 +0,0 @@
-# Single-node M3 Ultra benchmarks
-#
-# Shared constraints applied to ALL benchmarks in this file.
-constraints = [
-    "All(MacOsBuild(=25D125))",
-    "Hosts(=1)",
-    "All(Chip(m3_ultra))",
-    "All(GpuCores(=80))",
-]
-
-[topology]
-type = "none"
-
-# Default args merged into each benchmark's args (benchmark-level args win).
-[defaults]
-pp = [512, 2048, 8192, 16384]
-tg = 128
-
-[[benchmark]]
-model = "mlx-community/Meta-Llama-3.1-70B-Instruct-4bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/gpt-oss-120b-MXFP4-Q8"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/GLM-4.7-Flash-8bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-Coder-Next-6bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-30B-A3B-8bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-0.6B-4bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-0.6B-8bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Llama-3.2-1B-Instruct-4bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Llama-3.2-3B-Instruct-4bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Llama-3.2-3B-Instruct-8bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Meta-Llama-3.1-8B-Instruct-8bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/gpt-oss-20b-MXFP4-Q8"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-30B-A3B-4bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/GLM-4.7-Flash-4bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/GLM-4.7-Flash-5bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/GLM-4.7-Flash-6bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Llama-3.3-70B-Instruct-4bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-Coder-Next-4bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-Coder-Next-5bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-Coder-Next-8bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-Next-80B-A3B-Instruct-4bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-Next-80B-A3B-Instruct-8bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-Next-80B-A3B-Thinking-8bit"
-extra_constraints = ["All(Memory(>=96GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Llama-3.3-70B-Instruct-8bit"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/llama-3.3-70b-instruct-fp16"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/GLM-4.5-Air-8bit"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/GLM-4.5-Air-bf16"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/GLM-4.7-4bit"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/MiniMax-M2.1-3bit"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/MiniMax-M2.1-8bit"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-Coder-Next-bf16"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Step-3.5-Flash-4bit"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Step-3.5-Flash-6bit"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Step-3.5-Flash-8Bit"
-extra_constraints = ["All(Memory(>=256GiB))"]
-
-[[benchmark]]
-model = "mlx-community/DeepSeek-V3.1-4bit"
-extra_constraints = ["All(Memory(>=512GiB))"]
-
-[[benchmark]]
-model = "mlx-community/GLM-4.7-6bit"
-extra_constraints = ["All(Memory(>=512GiB))"]
-
-[[benchmark]]
-model = "mlx-community/GLM-4.7-8bit-gs32"
-extra_constraints = ["All(Memory(>=512GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-235B-A22B-Instruct-2507-8bit"
-extra_constraints = ["All(Memory(>=512GiB))"]
-
-[[benchmark]]
-model = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit"
-extra_constraints = ["All(Memory(>=512GiB))"]
@@ -14,7 +14,6 @@
    totalTokens,
    thinkingEnabled as thinkingEnabledStore,
    setConversationThinking,
-    stopGeneration,
  } from "$lib/stores/app.svelte";
  import ChatAttachments from "./ChatAttachments.svelte";
  import ImageParamsPanel from "./ImageParamsPanel.svelte";
@@ -104,7 +103,7 @@
  const modelSupportsThinking = $derived(() => {
    if (!currentModel) return false;
    const caps = modelCapabilities[currentModel] || [];
-    return caps.includes("thinking_toggle") && caps.includes("text");
+    return caps.includes("thinking") && caps.includes("text");
  });

  const isEditOnlyWithoutImage = $derived(
@@ -654,92 +653,86 @@
        style="min-height: 28px; max-height: 150px;"
      ></textarea>

-      {#if loading}
-        <button
-          type="button"
-          onclick={() => stopGeneration()}
-          class="px-2.5 sm:px-4 py-1.5 sm:py-2 rounded text-xs sm:text-xs tracking-[0.1em] sm:tracking-[0.15em] font-medium transition-all duration-200 whitespace-nowrap bg-exo-medium-gray/70 text-exo-light-gray hover:bg-exo-medium-gray hover:text-white"
-          aria-label="Stop generation"
-        >
+      <button
+        type="submit"
+        disabled={!canSend || loading || isEditOnlyWithoutImage}
+        class="px-2.5 sm:px-4 py-1.5 sm:py-2 rounded text-xs sm:text-xs tracking-[0.1em] sm:tracking-[0.15em] uppercase font-medium transition-all duration-200 whitespace-nowrap
+					{!canSend || loading || isEditOnlyWithoutImage
+          ? 'bg-exo-medium-gray/50 text-exo-light-gray cursor-not-allowed'
+          : 'bg-exo-yellow text-exo-black hover:bg-exo-yellow-darker hover:shadow-[0_0_20px_rgba(255,215,0,0.3)]'}"
+        aria-label={shouldShowEditMode
+          ? "Edit image"
+          : isImageModel()
+            ? "Generate image"
+            : "Send message"}
+      >
+        {#if loading}
          <span class="inline-flex items-center gap-1 sm:gap-2">
-            <svg
-              class="w-3 h-3 sm:w-3.5 sm:h-3.5"
-              fill="currentColor"
-              viewBox="0 0 24 24"
+            <span
+              class="w-2.5 h-2.5 sm:w-3 sm:h-3 border-2 border-current border-t-transparent rounded-full animate-spin"
+            ></span>
+            <span class="hidden sm:inline"
+              >{shouldShowEditMode
+                ? "EDITING"
+                : isImageModel()
+                  ? "GENERATING"
+                  : "PROCESSING"}</span
            >
-              <rect x="6" y="6" width="12" height="12" rx="1" />
-            </svg>
-            <span class="hidden sm:inline">Cancel</span>
+            <span class="sm:hidden">...</span>
          </span>
-        </button>
-      {:else}
-        <button
-          type="submit"
-          disabled={!canSend || isEditOnlyWithoutImage}
-          class="px-2.5 sm:px-4 py-1.5 sm:py-2 rounded text-xs sm:text-xs tracking-[0.1em] sm:tracking-[0.15em] uppercase font-medium transition-all duration-200 whitespace-nowrap
-					{!canSend || isEditOnlyWithoutImage
-            ? 'bg-exo-medium-gray/50 text-exo-light-gray cursor-not-allowed'
-            : 'bg-exo-yellow text-exo-black hover:bg-exo-yellow-darker hover:shadow-[0_0_20px_rgba(255,215,0,0.3)]'}"
-          aria-label={shouldShowEditMode
-            ? "Edit image"
-            : isImageModel()
-              ? "Generate image"
-              : "Send message"}
-        >
-          {#if shouldShowEditMode}
-            <span class="inline-flex items-center gap-1.5">
-              <svg
-                class="w-3.5 h-3.5"
-                fill="none"
-                viewBox="0 0 24 24"
-                stroke="currentColor"
-                stroke-width="2"
-              >
-                <path
-                  stroke-linecap="round"
-                  stroke-linejoin="round"
-                  d="M11 5H6a2 2 0 00-2 2v11a2 2 0 002 2h11a2 2 0 002-2v-5m-1.414-9.414a2 2 0 112.828 2.828L11.828 15H9v-2.828l8.586-8.586z"
-                />
-              </svg>
-              <span>EDIT</span>
-            </span>
-          {:else if isEditOnlyWithoutImage}
-            <span class="inline-flex items-center gap-1.5">
-              <svg
-                class="w-3.5 h-3.5"
-                fill="none"
-                viewBox="0 0 24 24"
-                stroke="currentColor"
-                stroke-width="2"
-              >
-                <path
-                  stroke-linecap="round"
-                  stroke-linejoin="round"
-                  d="M11 5H6a2 2 0 00-2 2v11a2 2 0 002 2h11a2 2 0 002-2v-5m-1.414-9.414a2 2 0 112.828 2.828L11.828 15H9v-2.828l8.586-8.586z"
-                />
-              </svg>
-              <span>EDIT</span>
-            </span>
-          {:else if isImageModel()}
-            <span class="inline-flex items-center gap-1.5">
-              <svg
-                class="w-3.5 h-3.5"
-                fill="none"
-                viewBox="0 0 24 24"
-                stroke="currentColor"
-                stroke-width="2"
-              >
-                <rect x="3" y="3" width="18" height="18" rx="2" ry="2" />
-                <circle cx="8.5" cy="8.5" r="1.5" />
-                <polyline points="21 15 16 10 5 21" />
-              </svg>
-              <span>GENERATE</span>
-            </span>
-          {:else}
-            SEND
-          {/if}
-        </button>
-      {/if}
+        {:else if shouldShowEditMode}
+          <span class="inline-flex items-center gap-1.5">
+            <svg
+              class="w-3.5 h-3.5"
+              fill="none"
+              viewBox="0 0 24 24"
+              stroke="currentColor"
+              stroke-width="2"
+            >
+              <path
+                stroke-linecap="round"
+                stroke-linejoin="round"
+                d="M11 5H6a2 2 0 00-2 2v11a2 2 0 002 2h11a2 2 0 002-2v-5m-1.414-9.414a2 2 0 112.828 2.828L11.828 15H9v-2.828l8.586-8.586z"
+              />
+            </svg>
+            <span>EDIT</span>
+          </span>
+        {:else if isEditOnlyWithoutImage}
+          <span class="inline-flex items-center gap-1.5">
+            <svg
+              class="w-3.5 h-3.5"
+              fill="none"
+              viewBox="0 0 24 24"
+              stroke="currentColor"
+              stroke-width="2"
+            >
+              <path
+                stroke-linecap="round"
+                stroke-linejoin="round"
+                d="M11 5H6a2 2 0 00-2 2v11a2 2 0 002 2h11a2 2 0 002-2v-5m-1.414-9.414a2 2 0 112.828 2.828L11.828 15H9v-2.828l8.586-8.586z"
+              />
+            </svg>
+            <span>EDIT</span>
+          </span>
+        {:else if isImageModel()}
+          <span class="inline-flex items-center gap-1.5">
+            <svg
+              class="w-3.5 h-3.5"
+              fill="none"
+              viewBox="0 0 24 24"
+              stroke="currentColor"
+              stroke-width="2"
+            >
+              <rect x="3" y="3" width="18" height="18" rx="2" ry="2" />
+              <circle cx="8.5" cy="8.5" r="1.5" />
+              <polyline points="21 15 16 10 5 21" />
+            </svg>
+            <span>GENERATE</span>
+          </span>
+        {:else}
+          SEND
+        {/if}
+      </button>
    </div>

    <!-- Bottom accent line -->
@@ -3,17 +3,16 @@
    messages,
    currentResponse,
    isLoading,
-    prefillProgress,
    deleteMessage,
    editAndRegenerate,
    regenerateLastResponse,
    regenerateFromToken,
    setEditingImage,
  } from "$lib/stores/app.svelte";
+  import type { Message } from "$lib/stores/app.svelte";
  import type { MessageAttachment } from "$lib/stores/app.svelte";
  import MarkdownContent from "./MarkdownContent.svelte";
  import TokenHeatmap from "./TokenHeatmap.svelte";
-  import PrefillProgressBar from "./PrefillProgressBar.svelte";
  import ImageLightbox from "./ImageLightbox.svelte";

  interface Props {
@@ -26,7 +25,6 @@
  const messageList = $derived(messages());
  const response = $derived(currentResponse());
  const loading = $derived(isLoading());
-  const prefill = $derived(prefillProgress());

  // Scroll management - user controls scroll, show button when not at bottom
  const SCROLL_THRESHOLD = 100;
@@ -430,9 +428,6 @@
            {:else}
              <!-- Assistant message styling -->
              <div class="p-3 sm:p-4">
-                {#if loading && isLastAssistantMessage(message.id) && prefill && !message.content}
-                  <PrefillProgressBar progress={prefill} class="mb-3" />
-                {/if}
                {#if message.thinking && message.thinking.trim().length > 0}
                  <div
                    class="mb-3 rounded border border-exo-yellow/20 bg-exo-black/40"
@@ -185,11 +185,7 @@

    let instanceType: string | null = null;
    if (instanceTag === "MlxRingInstance") instanceType = "MLX Ring";
-    else if (
-      instanceTag === "MlxIbvInstance" ||
-      instanceTag === "MlxJacclInstance"
-    )
-      instanceType = "MLX RDMA";
+    else if (instanceTag === "MlxJacclInstance") instanceType = "MLX RDMA";

    let sharding: string | null = null;
    const inst = instance as {
@@ -26,8 +26,7 @@
    downloadedOnNodes = [],
  }: HuggingFaceResultItemProps = $props();

-  function formatNumber(num: number | undefined): string {
-    if (num == null) return "0";
+  function formatNumber(num: number): string {
    if (num >= 1000000) {
      return `${(num / 1000000).toFixed(1)}M`;
    } else if (num >= 1000) {
@@ -59,14 +59,13 @@
  }

  const sizeOptions: ImageGenerationParams["size"][] = [
-    "auto",
    "512x512",
    "768x768",
    "1024x1024",
    "1024x768",
    "768x1024",
-    "1024x1536",
-    "1536x1024",
+    "1024x1365",
+    "1365x1024",
  ];

  const qualityOptions: ImageGenerationParams["quality"][] = [
@@ -177,90 +176,92 @@
 <div class="border-b border-exo-medium-gray/30 px-3 py-2">
  <!-- Basic params row -->
  <div class="flex items-center gap-3 flex-wrap">
-    <!-- Size -->
-    <div class="flex items-center gap-1.5">
-      <span class="text-xs text-exo-light-gray uppercase tracking-wider"
-        >SIZE:</span
-      >
-      <div class="relative">
-        <button
-          bind:this={sizeButtonRef}
-          type="button"
-          onclick={() => (isSizeDropdownOpen = !isSizeDropdownOpen)}
-          class="bg-exo-medium-gray/50 border border-exo-yellow/30 rounded pl-2 pr-6 py-1 text-xs font-mono text-exo-yellow cursor-pointer transition-all duration-200 hover:border-exo-yellow/50 focus:outline-none focus:border-exo-yellow/70 {isSizeDropdownOpen
-            ? 'border-exo-yellow/70'
-            : ''}"
+    <!-- Size (hidden in edit mode - output size comes from input image) -->
+    {#if !isEditMode}
+      <div class="flex items-center gap-1.5">
+        <span class="text-xs text-exo-light-gray uppercase tracking-wider"
+          >SIZE:</span
        >
-          {params.size.toUpperCase()}
-        </button>
-        <div
-          class="absolute right-1.5 top-1/2 -translate-y-1/2 pointer-events-none transition-transform duration-200 {isSizeDropdownOpen
-            ? 'rotate-180'
-            : ''}"
-        >
-          <svg
-            class="w-3 h-3 text-exo-yellow/60"
-            fill="none"
-            viewBox="0 0 24 24"
-            stroke="currentColor"
+        <div class="relative">
+          <button
+            bind:this={sizeButtonRef}
+            type="button"
+            onclick={() => (isSizeDropdownOpen = !isSizeDropdownOpen)}
+            class="bg-exo-medium-gray/50 border border-exo-yellow/30 rounded pl-2 pr-6 py-1 text-xs font-mono text-exo-yellow cursor-pointer transition-all duration-200 hover:border-exo-yellow/50 focus:outline-none focus:border-exo-yellow/70 {isSizeDropdownOpen
+              ? 'border-exo-yellow/70'
+              : ''}"
          >
-            <path
-              stroke-linecap="round"
-              stroke-linejoin="round"
-              stroke-width="2"
-              d="M19 9l-7 7-7-7"
-            />
-          </svg>
-        </div>
-      </div>
-
-      {#if isSizeDropdownOpen}
-        <!-- Backdrop to close dropdown -->
-        <button
-          type="button"
-          class="fixed inset-0 z-[9998] cursor-default"
-          onclick={() => (isSizeDropdownOpen = false)}
-          aria-label="Close dropdown"
-        ></button>
-
-        <!-- Dropdown Panel - fixed positioning to escape overflow:hidden -->
-        <div
-          class="fixed bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-[9999] max-h-48 overflow-y-auto overflow-x-hidden min-w-max"
-          style="bottom: calc(100vh - {sizeDropdownPosition()
-            .top}px + 4px); left: {sizeDropdownPosition().left}px;"
-        >
-          <div class="py-1">
-            {#each sizeOptions as size}
-              <button
-                type="button"
-                onclick={() => selectSize(size)}
-                class="w-full px-3 py-1.5 text-left text-xs font-mono tracking-wide transition-colors duration-100 flex items-center gap-2 {params.size ===
-                size
-                  ? 'bg-transparent text-exo-yellow'
-                  : 'text-exo-light-gray hover:text-exo-yellow'}"
-              >
-                {#if params.size === size}
-                  <svg
-                    class="w-3 h-3 flex-shrink-0"
-                    fill="currentColor"
-                    viewBox="0 0 20 20"
-                  >
-                    <path
-                      fill-rule="evenodd"
-                      d="M16.707 5.293a1 1 0 010 1.414l-8 8a1 1 0 01-1.414 0l-4-4a1 1 0 011.414-1.414L8 12.586l7.293-7.293a1 1 0 011.414 0z"
-                      clip-rule="evenodd"
-                    />
-                  </svg>
-                {:else}
-                  <span class="w-3"></span>
-                {/if}
-                <span>{size.toUpperCase()}</span>
-              </button>
-            {/each}
+            {params.size}
+          </button>
+          <div
+            class="absolute right-1.5 top-1/2 -translate-y-1/2 pointer-events-none transition-transform duration-200 {isSizeDropdownOpen
+              ? 'rotate-180'
+              : ''}"
+          >
+            <svg
+              class="w-3 h-3 text-exo-yellow/60"
+              fill="none"
+              viewBox="0 0 24 24"
+              stroke="currentColor"
+            >
+              <path
+                stroke-linecap="round"
+                stroke-linejoin="round"
+                stroke-width="2"
+                d="M19 9l-7 7-7-7"
+              />
+            </svg>
          </div>
        </div>
-      {/if}
-    </div>
+
+        {#if isSizeDropdownOpen}
+          <!-- Backdrop to close dropdown -->
+          <button
+            type="button"
+            class="fixed inset-0 z-[9998] cursor-default"
+            onclick={() => (isSizeDropdownOpen = false)}
+            aria-label="Close dropdown"
+          ></button>
+
+          <!-- Dropdown Panel - fixed positioning to escape overflow:hidden -->
+          <div
+            class="fixed bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-[9999] max-h-48 overflow-y-auto min-w-max"
+            style="bottom: calc(100vh - {sizeDropdownPosition()
+              .top}px + 4px); left: {sizeDropdownPosition().left}px;"
+          >
+            <div class="py-1">
+              {#each sizeOptions as size}
+                <button
+                  type="button"
+                  onclick={() => selectSize(size)}
+                  class="w-full px-3 py-1.5 text-left text-xs font-mono tracking-wide transition-colors duration-100 flex items-center gap-2 {params.size ===
+                  size
+                    ? 'bg-transparent text-exo-yellow'
+                    : 'text-exo-light-gray hover:text-exo-yellow'}"
+                >
+                  {#if params.size === size}
+                    <svg
+                      class="w-3 h-3 flex-shrink-0"
+                      fill="currentColor"
+                      viewBox="0 0 20 20"
+                    >
+                      <path
+                        fill-rule="evenodd"
+                        d="M16.707 5.293a1 1 0 010 1.414l-8 8a1 1 0 01-1.414 0l-4-4a1 1 0 011.414-1.414L8 12.586l7.293-7.293a1 1 0 011.414 0z"
+                        clip-rule="evenodd"
+                      />
+                    </svg>
+                  {:else}
+                    <span class="w-3"></span>
+                  {/if}
+                  <span>{size}</span>
+                </button>
+              {/each}
+            </div>
+          </div>
+        {/if}
+      </div>
+    {/if}

    <!-- Quality -->
    <div class="flex items-center gap-1.5">
@@ -310,7 +311,7 @@

        <!-- Dropdown Panel - fixed positioning to escape overflow:hidden -->
        <div
-          class="fixed bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-[9999] max-h-48 overflow-y-auto overflow-x-hidden min-w-max"
+          class="fixed bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-[9999] max-h-48 overflow-y-auto min-w-max"
          style="bottom: calc(100vh - {qualityDropdownPosition()
            .top}px + 4px); left: {qualityDropdownPosition().left}px;"
        >
@@ -21,7 +21,7 @@
    } | null;
    nodes?: Record<string, NodeInfo>;
    sharding?: "Pipeline" | "Tensor";
-    runtime?: "MlxRing" | "MlxIbv" | "MlxJaccl";
+    runtime?: "MlxRing" | "MlxJaccl";
    onLaunch?: () => void;
    tags?: string[];
    apiPreview?: PlacementPreview | null;
@@ -348,7 +348,7 @@
  // Debug mode state
  const isDebugMode = $derived(debugMode());
  const topology = $derived(topologyData());
-  const isRdma = $derived(runtime === "MlxIbv" || runtime === "MlxJaccl");
+  const isRdma = $derived(runtime === "MlxJaccl");

  // Get interface name for an IP from node data
  function getInterfaceForIp(nodeId: string, ip?: string): string | null {
@@ -575,7 +575,7 @@
      >
        {runtime === "MlxRing"
          ? "MLX Ring"
-          : runtime === "MlxIbv" || runtime === "MlxJaccl"
+          : runtime === "MlxJaccl"
            ? "MLX RDMA"
            : runtime}
      </span>
@@ -1,70 +0,0 @@
-<script lang="ts">
-  import type { PrefillProgress } from "$lib/stores/app.svelte";
-
-  interface Props {
-    progress: PrefillProgress;
-    class?: string;
-  }
-
-  let { progress, class: className = "" }: Props = $props();
-
-  const percentage = $derived(
-    progress.total > 0
-      ? Math.round((progress.processed / progress.total) * 100)
-      : 0,
-  );
-
-  const etaText = $derived.by(() => {
-    if (progress.processed <= 0 || progress.total <= 0) return null;
-    const elapsedMs = performance.now() - progress.startedAt;
-    if (elapsedMs < 200) return null; // need a minimum sample window
-    const tokensPerMs = progress.processed / elapsedMs;
-    const remainingTokens = progress.total - progress.processed;
-    const remainingMs = remainingTokens / tokensPerMs;
-    const remainingSec = Math.ceil(remainingMs / 1000);
-    if (remainingSec <= 0) return null;
-    if (remainingSec < 60) return `~${remainingSec}s remaining`;
-    const mins = Math.floor(remainingSec / 60);
-    const secs = remainingSec % 60;
-    return `~${mins}m ${secs}s remaining`;
-  });
-
-  function formatTokenCount(count: number | undefined): string {
-    if (count == null) return "0";
-    if (count >= 1000) {
-      return `${(count / 1000).toFixed(1)}k`;
-    }
-    return count.toString();
-  }
-</script>
-
-<div class="prefill-progress {className}">
-  <div
-    class="flex items-center justify-between text-xs text-exo-light-gray mb-1"
-  >
-    <span>Processing prompt</span>
-    <span class="font-mono">
-      {formatTokenCount(progress.processed)} / {formatTokenCount(
-        progress.total,
-      )} tokens
-    </span>
-  </div>
-  <div class="h-1.5 bg-exo-black/60 rounded-full overflow-hidden">
-    <div
-      class="h-full bg-exo-yellow rounded-full transition-all duration-150 ease-out"
-      style="width: {percentage}%"
-    ></div>
-  </div>
-  <div
-    class="flex items-center justify-between text-xs text-exo-light-gray/70 mt-0.5 font-mono"
-  >
-    <span>{etaText ?? ""}</span>
-    <span>{percentage}%</span>
-  </div>
-</div>
-
-<style>
-  .prefill-progress {
-    width: 100%;
-  }
-</style>
@@ -168,7 +168,7 @@ export interface ModelDownloadStatus {
 export interface PlacementPreview {
  model_id: string;
  sharding: "Pipeline" | "Tensor";
-  instance_meta: "MlxRing" | "MlxIbv" | "MlxJaccl";
+  instance_meta: "MlxRing" | "MlxJaccl";
  instance: unknown | null;
  memory_delta_by_node: Record<string, number> | null;
  error: string | null;
@@ -219,7 +219,6 @@ interface RawStateResponse {
    string,
    {
      MlxRingInstance?: Instance;
-      MlxIbvInstance?: Instance;
      MlxJacclInstance?: Instance;
    }
  >;
@@ -250,11 +249,20 @@ interface RawStateResponse {
  >;
  // Thunderbolt bridge cycles (nodes with bridge enabled forming loops)
  thunderboltBridgeCycles?: string[][];
-  // Disk usage per node
-  nodeDisk?: Record<
-    string,
-    { total: { inBytes: number }; available: { inBytes: number } }
-  >;
+  // MetaInstances (declarative instance constraints)
+  metaInstances?: Record<string, MetaInstanceData>;
+}
+
+export interface MetaInstanceData {
+  metaInstanceId: string;
+  modelId: string;
+  sharding: string;
+  instanceMeta: string;
+  minNodes: number;
+  nodeIds: string[] | null;
+  placementError: string | null;
+  consecutiveFailures: number;
+  lastFailureError: string | null;
 }

 export interface MessageAttachment {
@@ -278,13 +286,6 @@ export interface TokenData {
  topLogprobs: TopLogprob[];
 }

-export interface PrefillProgress {
-  processed: number;
-  total: number;
-  /** Timestamp (performance.now()) when prefill started. */
-  startedAt: number;
-}
-
 export interface Message {
  id: string;
  role: "user" | "assistant" | "system";
@@ -318,14 +319,13 @@ const IMAGE_PARAMS_STORAGE_KEY = "exo-image-generation-params";
 export interface ImageGenerationParams {
  // Basic params
  size:
-    | "auto"
    | "512x512"
    | "768x768"
    | "1024x1024"
    | "1024x768"
    | "768x1024"
-    | "1024x1536"
-    | "1536x1024";
+    | "1024x1365"
+    | "1365x1024";
  quality: "low" | "medium" | "high";
  outputFormat: "png" | "jpeg";
  numImages: number;
@@ -349,7 +349,7 @@ export interface EditingImage {
 }

 const DEFAULT_IMAGE_PARAMS: ImageGenerationParams = {
-  size: "auto",
+  size: "1024x1024",
  quality: "medium",
  outputFormat: "png",
  numImages: 1,
@@ -532,10 +532,6 @@ class AppStore {
  ttftMs = $state<number | null>(null); // Time to first token in ms
  tps = $state<number | null>(null); // Tokens per second
  totalTokens = $state<number>(0); // Total tokens in current response
-  prefillProgress = $state<PrefillProgress | null>(null);
-
-  // Abort controller for stopping generation
-  private currentAbortController: AbortController | null = null;

  // Topology state
  topologyData = $state<TopologyData | null>(null);
@@ -553,6 +549,7 @@ class AppStore {
  isLoadingPreviews = $state(false);
  previewNodeFilter = $state<Set<string>>(new Set());
  lastUpdate = $state<number | null>(null);
+  metaInstances = $state<Record<string, MetaInstanceData>>({});
  nodeIdentities = $state<Record<string, RawNodeIdentity>>({});
  thunderboltBridgeCycles = $state<string[][]>([]);
  nodeThunderbolt = $state<
@@ -912,11 +909,7 @@ class AppStore {

    let instanceType: string | null = null;
    if (instanceTag === "MlxRingInstance") instanceType = "MLX Ring";
-    else if (
-      instanceTag === "MlxIbvInstance" ||
-      instanceTag === "MlxJacclInstance"
-    )
-      instanceType = "MLX RDMA";
+    else if (instanceTag === "MlxJacclInstance") instanceType = "MLX RDMA";

    let sharding: string | null = null;
    const inst = instance as {
@@ -1281,6 +1274,8 @@ class AppStore {
      if (data.downloads) {
        this.downloads = data.downloads;
      }
+      // MetaInstances
+      this.metaInstances = data.metaInstances ?? {};
      if (data.nodeDisk) {
        this.nodeDisk = data.nodeDisk;
      }
@@ -1659,12 +1654,11 @@ class AppStore {
      if (!reader) throw new Error("No response body");

      let fullContent = prefixText;
-      let streamedThinking = "";
      const collectedTokens: TokenData[] = [...tokensToKeep];

      interface ChatCompletionChunk {
        choices?: Array<{
-          delta?: { content?: string; reasoning_content?: string };
+          delta?: { content?: string };
          logprobs?: {
            content?: Array<{
              token: string;
@@ -1685,7 +1679,6 @@ class AppStore {
        (parsed) => {
          const choice = parsed.choices?.[0];
          const delta = choice?.delta?.content;
-          const thinkingDelta = choice?.delta?.reasoning_content;

          // Collect logprobs data
          const logprobsContent = choice?.logprobs?.content;
@@ -1704,11 +1697,7 @@ class AppStore {
            }
          }

-          if (thinkingDelta) {
-            streamedThinking += thinkingDelta;
-          }
-
-          if (delta || thinkingDelta) {
+          if (delta) {
            if (firstTokenTime === null) {
              firstTokenTime = performance.now();
              this.ttftMs = firstTokenTime - requestStartTime;
@@ -1722,14 +1711,9 @@ class AppStore {
              this.tps = ((tokenCount - tokensToKeep.length) / elapsed) * 1000;
            }

-            if (delta) {
-              fullContent += delta;
-            }
-            const { displayContent, thinkingContent: tagThinking } =
+            fullContent += delta;
+            const { displayContent, thinkingContent } =
              this.stripThinkingTags(fullContent);
-            const combinedThinking = [streamedThinking, tagThinking]
-              .filter(Boolean)
-              .join("\n\n");

            if (this.activeConversationId === targetConversationId) {
              this.currentResponse = displayContent;
@@ -1741,7 +1725,7 @@ class AppStore {
              messageId,
              (m) => {
                m.content = displayContent;
-                m.thinking = combinedThinking || undefined;
+                m.thinking = thinkingContent || undefined;
                m.tokens = [...collectedTokens];
              },
            );
@@ -1753,14 +1737,11 @@ class AppStore {

      // Final update
      if (this.conversationExists(targetConversationId)) {
-        const { displayContent, thinkingContent: tagThinking } =
+        const { displayContent, thinkingContent } =
          this.stripThinkingTags(fullContent);
-        const finalThinking = [streamedThinking, tagThinking]
-          .filter(Boolean)
-          .join("\n\n");
        this.updateConversationMessage(targetConversationId, messageId, (m) => {
          m.content = displayContent;
-          m.thinking = finalThinking || undefined;
+          m.thinking = thinkingContent || undefined;
          m.tokens = [...collectedTokens];
          if (this.ttftMs !== null) m.ttftMs = this.ttftMs;
          if (this.tps !== null) m.tps = this.tps;
@@ -1868,12 +1849,11 @@ class AppStore {
      }

      let streamedContent = "";
-      let streamedThinking = "";
      const collectedTokens: TokenData[] = [];

      interface ChatCompletionChunk {
        choices?: Array<{
-          delta?: { content?: string; reasoning_content?: string };
+          delta?: { content?: string };
          logprobs?: {
            content?: Array<{
              token: string;
@@ -1894,7 +1874,6 @@ class AppStore {
        (parsed) => {
          const choice = parsed.choices?.[0];
          const delta = choice?.delta?.content;
-          const thinkingDelta = choice?.delta?.reasoning_content;

          // Collect logprobs data
          const logprobsContent = choice?.logprobs?.content;
@@ -1913,19 +1892,10 @@ class AppStore {
            }
          }

-          if (thinkingDelta) {
-            streamedThinking += thinkingDelta;
-          }
-
-          if (delta || thinkingDelta) {
-            if (delta) {
-              streamedContent += delta;
-            }
-            const { displayContent, thinkingContent: tagThinking } =
+          if (delta) {
+            streamedContent += delta;
+            const { displayContent, thinkingContent } =
              this.stripThinkingTags(streamedContent);
-            const combinedThinking = [streamedThinking, tagThinking]
-              .filter(Boolean)
-              .join("\n\n");

            // Only update currentResponse if target conversation is active
            if (this.activeConversationId === targetConversationId) {
@@ -1938,7 +1908,7 @@ class AppStore {
              assistantMessage.id,
              (msg) => {
                msg.content = displayContent;
-                msg.thinking = combinedThinking || undefined;
+                msg.thinking = thinkingContent || undefined;
                msg.tokens = [...collectedTokens];
              },
            );
@@ -1950,17 +1920,14 @@ class AppStore {

      // Final cleanup of the message (if conversation still exists)
      if (this.conversationExists(targetConversationId)) {
-        const { displayContent, thinkingContent: tagThinking } =
+        const { displayContent, thinkingContent } =
          this.stripThinkingTags(streamedContent);
-        const finalThinking = [streamedThinking, tagThinking]
-          .filter(Boolean)
-          .join("\n\n");
        this.updateConversationMessage(
          targetConversationId,
          assistantMessage.id,
          (msg) => {
            msg.content = displayContent;
-            msg.thinking = finalThinking || undefined;
+            msg.thinking = thinkingContent || undefined;
            msg.tokens = [...collectedTokens];
          },
        );
@@ -2049,7 +2016,6 @@ class AppStore {
    reader: ReadableStreamDefaultReader<Uint8Array>,
    targetConversationId: string,
    onChunk: (parsed: T) => void,
-    onEvent?: Record<string, (data: unknown) => void>,
  ): Promise<void> {
    const decoder = new TextDecoder();
    let buffer = "";
@@ -2070,24 +2036,6 @@ class AppStore {
        const trimmed = line.trim();
        if (!trimmed) continue;

-        // Handle SSE comments (": key json") for prefill progress etc.
-        if (trimmed.startsWith(": ") && onEvent) {
-          const comment = trimmed.slice(2);
-          const spaceIdx = comment.indexOf(" ");
-          if (spaceIdx > 0) {
-            const key = comment.slice(0, spaceIdx);
-            if (onEvent[key]) {
-              try {
-                const parsed = JSON.parse(comment.slice(spaceIdx + 1));
-                onEvent[key](parsed);
-              } catch {
-                // Skip malformed JSON in comment
-              }
-            }
-          }
-          continue;
-        }
-
        if (trimmed.startsWith("data: ")) {
          const data = trimmed.slice(6);
          if (data === "[DONE]") continue;
@@ -2319,9 +2267,6 @@ class AppStore {
      let firstTokenTime: number | null = null;
      let tokenCount = 0;

-      const abortController = new AbortController();
-      this.currentAbortController = abortController;
-
      const response = await fetch("/v1/chat/completions", {
        method: "POST",
        headers: {
@@ -2338,7 +2283,6 @@ class AppStore {
            enable_thinking: enableThinking,
          }),
        }),
-        signal: abortController.signal,
      });

      if (!response.ok) {
@@ -2352,11 +2296,10 @@ class AppStore {
      }

      let streamedContent = "";
-      let streamedThinking = "";

      interface ChatCompletionChunk {
        choices?: Array<{
-          delta?: { content?: string; reasoning_content?: string };
+          delta?: { content?: string };
          logprobs?: {
            content?: Array<{
              token: string;
@@ -2377,14 +2320,8 @@ class AppStore {
        reader,
        targetConversationId,
        (parsed) => {
-          // Clear prefill progress when first token data arrives
-          if (this.prefillProgress) {
-            this.prefillProgress = null;
-          }
-
          const choice = parsed.choices?.[0];
          const tokenContent = choice?.delta?.content;
-          const thinkingContent = choice?.delta?.reasoning_content;

          // Collect logprobs data
          const logprobsContent = choice?.logprobs?.content;
@@ -2403,11 +2340,7 @@ class AppStore {
            }
          }

-          if (thinkingContent) {
-            streamedThinking += thinkingContent;
-          }
-
-          if (tokenContent || thinkingContent) {
+          if (tokenContent) {
            // Track first token for TTFT
            if (firstTokenTime === null) {
              firstTokenTime = performance.now();
@@ -2424,16 +2357,11 @@ class AppStore {
              this.tps = (tokenCount / elapsed) * 1000;
            }

-            if (tokenContent) {
-              streamedContent += tokenContent;
-            }
+            streamedContent += tokenContent;

-            // Use stripThinkingTags as fallback for any <think> tags still in content
-            const { displayContent, thinkingContent: tagThinking } =
+            // Strip thinking tags for display and extract thinking content
+            const { displayContent, thinkingContent } =
              this.stripThinkingTags(streamedContent);
-            const combinedThinking = [streamedThinking, tagThinking]
-              .filter(Boolean)
-              .join("\n\n");

            // Only update currentResponse if target conversation is active
            if (this.activeConversationId === targetConversationId) {
@@ -2446,7 +2374,7 @@ class AppStore {
              assistantMessage.id,
              (msg) => {
                msg.content = displayContent;
-                msg.thinking = combinedThinking || undefined;
+                msg.thinking = thinkingContent || undefined;
                msg.tokens = [...collectedTokens];
              },
            );
@@ -2454,27 +2382,8 @@ class AppStore {
            this.persistConversation(targetConversationId);
          }
        },
-        {
-          prefill_progress: (data) => {
-            // TaggedModel wraps as {"PrefillProgressChunk": {...}}
-            // model_dump_json() uses snake_case (by_alias defaults to False)
-            const raw = data as Record<string, unknown>;
-            const inner = (raw["PrefillProgressChunk"] ?? raw) as {
-              processed_tokens: number;
-              total_tokens: number;
-            };
-            this.prefillProgress = {
-              processed: inner.processed_tokens,
-              total: inner.total_tokens,
-              startedAt: this.prefillProgress?.startedAt ?? performance.now(),
-            };
-          },
-        },
      );

-      // Clear prefill progress after stream ends
-      this.prefillProgress = null;
-
      // Calculate final TPS
      if (firstTokenTime !== null && tokenCount > 1) {
        const totalGenerationTime = performance.now() - firstTokenTime;
@@ -2483,17 +2392,14 @@ class AppStore {

      // Final cleanup of the message (if conversation still exists)
      if (this.conversationExists(targetConversationId)) {
-        const { displayContent, thinkingContent: tagThinking } =
+        const { displayContent, thinkingContent } =
          this.stripThinkingTags(streamedContent);
-        const finalThinking = [streamedThinking, tagThinking]
-          .filter(Boolean)
-          .join("\n\n");
        this.updateConversationMessage(
          targetConversationId,
          assistantMessage.id,
          (msg) => {
            msg.content = displayContent;
-            msg.thinking = finalThinking || undefined;
+            msg.thinking = thinkingContent || undefined;
            msg.tokens = [...collectedTokens];
            // Store performance metrics on the message
            if (this.ttftMs !== null) {
@@ -2508,31 +2414,20 @@ class AppStore {
        this.persistConversation(targetConversationId);
      }
    } catch (error) {
-      if (error instanceof DOMException && error.name === "AbortError") {
-        // User stopped generation — not an error
-      } else {
-        console.error("Error sending message:", error);
-        this.handleStreamingError(
-          error,
-          targetConversationId,
-          assistantMessage.id,
-          "Failed to get response",
-        );
-      }
+      console.error("Error sending message:", error);
+      this.handleStreamingError(
+        error,
+        targetConversationId,
+        assistantMessage.id,
+        "Failed to get response",
+      );
    } finally {
-      this.currentAbortController = null;
-      this.prefillProgress = null;
      this.isLoading = false;
      this.currentResponse = "";
      this.saveConversationsToStorage();
    }
  }

-  stopGeneration(): void {
-    this.currentAbortController?.abort();
-    this.currentAbortController = null;
-  }
-
  /**
   * Generate an image using the image generation API
   */
@@ -3159,9 +3054,9 @@ export const isLoading = () => appStore.isLoading;
 export const ttftMs = () => appStore.ttftMs;
 export const tps = () => appStore.tps;
 export const totalTokens = () => appStore.totalTokens;
-export const prefillProgress = () => appStore.prefillProgress;
 export const topologyData = () => appStore.topologyData;
 export const instances = () => appStore.instances;
+export const metaInstances = () => appStore.metaInstances;
 export const runners = () => appStore.runners;
 export const downloads = () => appStore.downloads;
 export const nodeDisk = () => appStore.nodeDisk;
@@ -3177,7 +3072,6 @@ export const topologyOnlyMode = () => appStore.getTopologyOnlyMode();
 export const chatSidebarVisible = () => appStore.getChatSidebarVisible();

 // Actions
-export const stopGeneration = () => appStore.stopGeneration();
 export const startChat = () => appStore.startChat();
 export const sendMessage = (
  content: string,
@@ -74,6 +74,7 @@
      perSystem =
        { config, self', inputs', pkgs, lib, system, ... }:
        let
+          fenixToolchain = inputs'.fenix.packages.complete;
          # Use pinned nixpkgs for swift-format (swift is broken on x86_64-linux in newer nixpkgs)
          pkgsSwift = import inputs.nixpkgs-swift { inherit system; };
        in
@@ -114,7 +115,7 @@
          packages = lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin (
            let
              uvLock = builtins.fromTOML (builtins.readFile ./uv.lock);
-              mlxPackage = builtins.head (builtins.filter (p: p.name == "mlx" && p.source ? git) uvLock.package);
+              mlxPackage = builtins.head (builtins.filter (p: p.name == "mlx") uvLock.package);
              uvLockMlxVersion = mlxPackage.version;
            in
            {
@@ -41,16 +41,16 @@ let

  mlx = stdenv.mkDerivation rec {
    pname = "mlx";
-    version = let v = "0.30.7.dev20260220+db176851"; 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;

    src = fetchFromGitHub {
-      owner = "exo-explore";
+      owner = "ml-explore";
      repo = "mlx";
-      rev = "db176851b031631b414890d6bd31d3dfb2e82d25";
-      hash = "sha256-/76lKDMmAjs7Qqh9V0oIBoNPq1kv819JZ+UbVFF8pd4=";
+      tag = "v${version}";
+      hash = "sha256-avD5EGhwgmPdXLAyQSqTO6AXk/W3ziH+f6AetjK3Sdo=";
    };

    patches = [
@@ -17,9 +17,9 @@ dependencies = [
    "loguru>=0.7.3",
    "exo_pyo3_bindings", # rust bindings
    "anyio==4.11.0",
-    "mlx; sys_platform == 'darwin'",
+    "mlx==0.30.6; sys_platform == 'darwin'",
    "mlx[cpu]==0.30.6; sys_platform == 'linux'",
-    "mlx-lm==0.30.7",
+    "mlx-lm==0.30.6",
    "tiktoken>=0.12.0", # required for kimi k2 tokenizer
    "hypercorn>=0.18.0",
    "openai-harmony>=0.0.8",
@@ -64,7 +64,6 @@ members = [

 [tool.uv.sources]
 exo_pyo3_bindings = { workspace = true }
-mlx = { git = "https://github.com/exo-explore/mlx.git", branch = "pipe-sidechannel", marker = "sys_platform == 'darwin'" }
 #mlx-lm = { git = "https://github.com/davidmcc73/mlx-lm", branch = "stable" }
 # Uncomment to use local mlx/mlx-lm development versions:
 # mlx = { path = "/Users/Shared/mlx", editable=true }
@@ -58,21 +58,6 @@
        lib.optionalAttrs pkgs.stdenv.hostPlatform.isLinux (
          (lib.mapAttrs (_: ignoreMissing) nvidiaPackages) // {
            mlx = ignoreMissing prev.mlx;
-            mlx-cuda-13 = prev.mlx-cuda-13.overrideAttrs (old: {
-              buildInputs = (old.buildInputs or [ ]) ++ [
-                final.nvidia-cublas
-                final.nvidia-cuda-nvrtc
-                final.nvidia-cudnn-cu13
-                final.nvidia-nccl-cu13
-              ];
-              preFixup = ''
-                addAutoPatchelfSearchPath ${final.nvidia-cublas}
-                addAutoPatchelfSearchPath ${final.nvidia-cuda-nvrtc}
-                addAutoPatchelfSearchPath ${final.nvidia-cudnn-cu13}
-                addAutoPatchelfSearchPath ${final.nvidia-nccl-cu13}
-              '';
-              autoPatchelfIgnoreMissingDeps = [ "libcuda.so.1" ];
-            });
            torch = ignoreMissing prev.torch;
            triton = ignoreMissing prev.triton;
          }
@@ -89,25 +74,14 @@
          linuxOverlay
        ]
      );
-      # mlx-cpu and mlx-cuda-13 both ship mlx/ site-packages files; keep first.
-      # mlx-cpu/mlx-cuda-13 and nvidia-cudnn-cu12/cu13 ship overlapping files.
-      venvCollisionPaths = lib.optionals pkgs.stdenv.hostPlatform.isLinux [
-        "lib/python3.13/site-packages/mlx*"
-        "lib/python3.13/site-packages/nvidia*"
-      ];
-
-      exoVenv = (pythonSet.mkVirtualEnv "exo-env" workspace.deps.default).overrideAttrs {
-        venvIgnoreCollisions = venvCollisionPaths;
-      };
+      exoVenv = pythonSet.mkVirtualEnv "exo-env" workspace.deps.default;

      # Virtual environment with dev dependencies for testing
-      testVenv = (pythonSet.mkVirtualEnv "exo-test-env" (
+      testVenv = pythonSet.mkVirtualEnv "exo-test-env" (
        workspace.deps.default // {
          exo = [ "dev" ]; # Include pytest, pytest-asyncio, pytest-env
        }
-      )).overrideAttrs {
-        venvIgnoreCollisions = venvCollisionPaths;
-      };
+      );

      mkPythonScript = name: path: pkgs.writeShellApplication {
        inherit name;
@@ -158,7 +132,6 @@
          exo-test-env = testVenv;
        } // {
        exo-bench = mkBenchScript "exo-bench" (inputs.self + /bench/exo_bench.py);
-        exo-eval-tool-calls = mkBenchScript "exo-eval-tool-calls" (inputs.self + /bench/eval_tool_calls.py);
        exo-get-all-models-on-cluster = mkSimplePythonScript "exo-get-all-models-on-cluster" (inputs.self + /tests/get_all_models_on_cluster.py);
      };

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "deepseek"
 quantization = "4bit"
 base_model = "DeepSeek V3.1"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 405874409472
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "deepseek"
 quantization = "8bit"
 base_model = "DeepSeek V3.1"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 765577920512
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "8bit"
 base_model = "GLM 4.5 Air"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 122406567936
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "bf16"
 base_model = "GLM 4.5 Air"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 229780750336
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "4bit"
 base_model = "GLM 4.7"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 198556925568
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "6bit"
 base_model = "GLM 4.7"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 286737579648
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "8bit"
 base_model = "GLM 4.7"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 396963397248
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "4bit"
 base_model = "GLM 4.7 Flash"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 19327352832
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "5bit"
 base_model = "GLM 4.7 Flash"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 22548578304
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "6bit"
 base_model = "GLM 4.7 Flash"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 26843545600
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "8bit"
 base_model = "GLM 4.7 Flash"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 34359738368
@@ -1,12 +0,0 @@
-model_id = "mlx-community/GLM-5-8bit-MXFP8"
-n_layers = 78
-hidden_size = 6144
-supports_tensor = true
-tasks = ["TextGeneration"]
-family = "glm"
-quantization = "8bit"
-base_model = "GLM-5"
-capabilities = ["text", "thinking"]
-
-[storage_size]
-in_bytes = 790517400864
@@ -1,12 +0,0 @@
-model_id = "mlx-community/GLM-5-MXFP4-Q8"
-n_layers = 78
-hidden_size = 6144
-supports_tensor = true
-tasks = ["TextGeneration"]
-family = "glm"
-quantization = "MXFP4-Q8"
-base_model = "GLM-5"
-capabilities = ["text", "thinking"]
-
-[storage_size]
-in_bytes = 405478939008
@@ -1,12 +0,0 @@
-model_id = "mlx-community/GLM-5"
-n_layers = 78
-hidden_size = 6144
-supports_tensor = true
-tasks = ["TextGeneration"]
-family = "glm"
-quantization = "bf16"
-base_model = "GLM-5"
-capabilities = ["text", "thinking"]
-
-[storage_size]
-in_bytes = 1487822475264
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "kimi"
 quantization = ""
 base_model = "Kimi K2"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 706522120192
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "kimi"
 quantization = ""
 base_model = "Kimi K2.5"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 662498705408
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "minimax"
 quantization = "3bit"
 base_model = "MiniMax M2.1"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 100086644736
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "minimax"
 quantization = "8bit"
 base_model = "MiniMax M2.1"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 242986745856
@@ -1,12 +0,0 @@
-model_id = "mlx-community/MiniMax-M2.5-4bit"
-n_layers = 62
-hidden_size = 3072
-supports_tensor = true
-tasks = ["TextGeneration"]
-family = "minimax"
-quantization = "4bit"
-base_model = "MiniMax M2.5"
-capabilities = ["text", "thinking"]
-
-[storage_size]
-in_bytes = 128666664960
@@ -1,12 +0,0 @@
-model_id = "mlx-community/MiniMax-M2.5-6bit"
-n_layers = 62
-hidden_size = 3072
-supports_tensor = true
-tasks = ["TextGeneration"]
-family = "minimax"
-quantization = "6bit"
-base_model = "MiniMax M2.5"
-capabilities = ["text", "thinking"]
-
-[storage_size]
-in_bytes = 185826705408
@@ -1,12 +0,0 @@
-model_id = "mlx-community/MiniMax-M2.5-8bit"
-n_layers = 62
-hidden_size = 3072
-supports_tensor = true
-tasks = ["TextGeneration"]
-family = "minimax"
-quantization = "8bit"
-base_model = "MiniMax M2.5"
-capabilities = ["text", "thinking"]
-
-[storage_size]
-in_bytes = 242986745856
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "4bit"
 base_model = "Qwen3 0.6B"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 342884352
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "8bit"
 base_model = "Qwen3 0.6B"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 698351616
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "4bit"
 base_model = "Qwen3 235B"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 141733920768
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "8bit"
 base_model = "Qwen3 235B"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 268435456000
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "4bit"
 base_model = "Qwen3 30B"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 17612931072
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "8bit"
 base_model = "Qwen3 30B"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 33279705088
@@ -3,10 +3,6 @@ n_layers = 48
 hidden_size = 2048
 supports_tensor = true
 tasks = ["TextGeneration"]
-family = "qwen"
-quantization = "4bit"
-base_model = "Qwen3 Coder Next"
-capabilities = ["text", "code"]

 [storage_size]
 in_bytes = 45644286500
@@ -3,10 +3,6 @@ n_layers = 48
 hidden_size = 2048
 supports_tensor = true
 tasks = ["TextGeneration"]
-family = "qwen"
-quantization = "5bit"
-base_model = "Qwen3 Coder Next"
-capabilities = ["text", "code"]

 [storage_size]
 in_bytes = 57657697020
@@ -3,10 +3,6 @@ n_layers = 48
 hidden_size = 2048
 supports_tensor = true
 tasks = ["TextGeneration"]
-family = "qwen"
-quantization = "6bit"
-base_model = "Qwen3 Coder Next"
-capabilities = ["text", "code"]

 [storage_size]
 in_bytes = 68899327465
@@ -3,10 +3,6 @@ n_layers = 48
 hidden_size = 2048
 supports_tensor = true
 tasks = ["TextGeneration"]
-family = "qwen"
-quantization = "8bit"
-base_model = "Qwen3 Coder Next"
-capabilities = ["text", "code"]

 [storage_size]
 in_bytes = 89357758772
@@ -3,10 +3,6 @@ n_layers = 48
 hidden_size = 2048
 supports_tensor = true
 tasks = ["TextGeneration"]
-family = "qwen"
-quantization = "bf16"
-base_model = "Qwen3 Coder Next"
-capabilities = ["text", "code"]

 [storage_size]
 in_bytes = 157548627945
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "4bit"
 base_model = "Qwen3 Next 80B"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 47080074240
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "8bit"
 base_model = "Qwen3 Next 80B"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 88814387200
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "step"
 quantization = "4bit"
 base_model = "Step 3.5 Flash"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 114572190076
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "step"
 quantization = "6bit"
 base_model = "Step 3.5 Flash"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 159039627774
@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "step"
 quantization = "8bit"
 base_model = "Step 3.5 Flash"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking"]

 [storage_size]
 in_bytes = 209082699847
@@ -0,0 +1,2 @@
+# we can manually exclude false-positive lint errors for dual packages (if in dependencies)
+#allowed-duplicate-crates = ["hashbrown"]
@@ -25,17 +25,17 @@ workspace = true
 networking = { workspace = true }

 # interop
-pyo3 = { version = "0.27.2", features = [
-    # "abi3-py313", # tells pyo3 (and maturin) to build using the stable ABI with minimum Python version 3.13
-    # "nightly", # enables better-supported GIL integration
+pyo3 = { version = "0.27.1", features = [
+    # "abi3-py311", # tells pyo3 (and maturin) to build using the stable ABI with minimum Python version 3.11
+    "nightly", # enables better-supported GIL integration
    "experimental-async", # async support in #[pyfunction] & #[pymethods]
    #"experimental-inspect", # inspection of generated binary => easier to automate type-hint generation
    #"py-clone", # adding Clone-ing of `Py<T>` without GIL (may cause panics - remove if panics happen)
-    # "multiple-pymethods", # allows multiple #[pymethods] sections per class
+    "multiple-pymethods", # allows multiple #[pymethods] sections per class

    # integrations with other libraries
-    # "arc_lock", "bigdecimal", "either", "hashbrown", "indexmap", "num-bigint", "num-complex", "num-rational",
-    # "ordered-float", "rust_decimal", "smallvec",
+    "arc_lock", "bigdecimal", "either", "hashbrown", "indexmap", "num-bigint", "num-complex", "num-rational",
+    "ordered-float", "rust_decimal", "smallvec",
    # "anyhow", "chrono", "chrono-local", "chrono-tz", "eyre", "jiff-02", "lock_api", "parking-lot", "time",  "serde",
 ] }
 pyo3-stub-gen = { version = "0.17.2" }
@@ -45,18 +45,33 @@ pyo3-log = "0.13.2"
 # macro dependencies
 extend = { workspace = true }
 delegate = { workspace = true }
+impl-trait-for-tuples = { workspace = true }
+derive_more = { workspace = true }
+pin-project = { workspace = true }

 # async runtime
 tokio = { workspace = true, features = ["full", "tracing"] }
-futures-lite = { workspace = true }
+futures = { workspace = true }

 # utility dependencies
+once_cell = "1.21.3"
+thread_local = "1.1.9"
 util = { workspace = true }
+thiserror = { workspace = true }
+#internment = { workspace = true }
+#recursion = { workspace = true }
+#generativity = { workspace = true }
+#itertools = { workspace = true }
+

 # Tracing
+#tracing = "0.1"
+#tracing-subscriber = "0.3"
+#console-subscriber = "0.1.5"
+#tracing-log = "0.2.0"
 log = { workspace = true }
 env_logger = "0.11"

+
 # Networking
 libp2p = { workspace = true, features = ["full"] }
-pin-project = "1.1.10"
@@ -19,7 +19,7 @@ class ConnectionUpdate:
        Whether this is a connection or disconnection event
        """
    @property
-    def peer_id(self) -> builtins.str:
+    def peer_id(self) -> PeerId:
        r"""
        Identity of the peer that we have connected to or disconnected from.
        """
@@ -40,27 +40,97 @@ class Keypair:
    Identity keypair of a node.
    """
    @staticmethod
-    def generate() -> Keypair:
+    def generate_ed25519() -> Keypair:
        r"""
        Generate a new Ed25519 keypair.
        """
    @staticmethod
-    def from_bytes(bytes: bytes) -> Keypair:
+    def generate_ecdsa() -> Keypair:
        r"""
-        Construct an Ed25519 keypair from secret key bytes
+        Generate a new ECDSA keypair.
+        """
+    @staticmethod
+    def generate_secp256k1() -> Keypair:
+        r"""
+        Generate a new Secp256k1 keypair.
+        """
+    @staticmethod
+    def from_protobuf_encoding(bytes: bytes) -> Keypair:
+        r"""
+        Decode a private key from a protobuf structure and parse it as a `Keypair`.
+        """
+    @staticmethod
+    def rsa_from_pkcs8(bytes: bytes) -> Keypair:
+        r"""
+        Decode an keypair from a DER-encoded secret key in PKCS#8 `PrivateKeyInfo`
+        format (i.e. unencrypted) as defined in [RFC5208].
+        
+        [RFC5208]: https://tools.ietf.org/html/rfc5208#section-5
+        """
+    @staticmethod
+    def secp256k1_from_der(bytes: bytes) -> Keypair:
+        r"""
+        Decode a keypair from a DER-encoded Secp256k1 secret key in an `ECPrivateKey`
+        structure as defined in [RFC5915].
+        
+        [RFC5915]: https://tools.ietf.org/html/rfc5915
+        """
+    @staticmethod
+    def ed25519_from_bytes(bytes: bytes) -> Keypair: ...
+    def to_protobuf_encoding(self) -> bytes:
+        r"""
+        Encode a private key as protobuf structure.
+        """
+    def to_peer_id(self) -> PeerId:
+        r"""
+        Convert the `Keypair` into the corresponding `PeerId`.
+        """
+
+@typing.final
+class Multiaddr:
+    r"""
+    Representation of a Multiaddr.
+    """
+    @staticmethod
+    def empty() -> Multiaddr:
+        r"""
+        Create a new, empty multiaddress.
+        """
+    @staticmethod
+    def with_capacity(n: builtins.int) -> Multiaddr:
+        r"""
+        Create a new, empty multiaddress with the given capacity.
+        """
+    @staticmethod
+    def from_bytes(bytes: bytes) -> Multiaddr:
+        r"""
+        Parse a `Multiaddr` value from its byte slice representation.
+        """
+    @staticmethod
+    def from_string(string: builtins.str) -> Multiaddr:
+        r"""
+        Parse a `Multiaddr` value from its string representation.
+        """
+    def len(self) -> builtins.int:
+        r"""
+        Return the length in bytes of this multiaddress.
+        """
+    def is_empty(self) -> builtins.bool:
+        r"""
+        Returns true if the length of this multiaddress is 0.
        """
    def to_bytes(self) -> bytes:
        r"""
-        Get the secret key bytes underlying the keypair
+        Return a copy of this [`Multiaddr`]'s byte representation.
        """
-    def to_node_id(self) -> builtins.str:
+    def to_string(self) -> builtins.str:
        r"""
-        Convert the `Keypair` into the corresponding `PeerId` string, which we use as our `NodeId`.
+        Convert a Multiaddr to a string.
        """

@typing.final
 class NetworkingHandle:
-    def __new__(cls, identity: Keypair) -> NetworkingHandle: ...
+    def __new__(cls, identity: Keypair, bootstrap_peers: builtins.list[builtins.str] = ...) -> NetworkingHandle: ...
    async def connection_update_recv(self) -> ConnectionUpdate:
        r"""
        Receives the next `ConnectionUpdate` from networking.
@@ -110,6 +180,37 @@ class NoPeersSubscribedToTopicError(builtins.Exception):
    def __repr__(self) -> builtins.str: ...
    def __str__(self) -> builtins.str: ...

+@typing.final
+class PeerId:
+    r"""
+    Identifier of a peer of the network.
+    
+    The data is a `CIDv0` compatible multihash of the protobuf encoded public key of the peer
+    as specified in [specs/peer-ids](https://github.com/libp2p/specs/blob/master/peer-ids/peer-ids.md).
+    """
+    @staticmethod
+    def random() -> PeerId:
+        r"""
+        Generates a random peer ID from a cryptographically secure PRNG.
+        
+        This is useful for randomly walking on a DHT, or for testing purposes.
+        """
+    @staticmethod
+    def from_bytes(bytes: bytes) -> PeerId:
+        r"""
+        Parses a `PeerId` from bytes.
+        """
+    def to_bytes(self) -> bytes:
+        r"""
+        Returns a raw bytes representation of this `PeerId`.
+        """
+    def to_base58(self) -> builtins.str:
+        r"""
+        Returns a base-58 encoded string of this `PeerId`.
+        """
+    def __repr__(self) -> builtins.str: ...
+    def __str__(self) -> builtins.str: ...
+
@typing.final
 class ConnectionUpdateType(enum.Enum):
    r"""
@@ -2,10 +2,11 @@
 //!

 use pin_project::pin_project;
+use pyo3::marker::Ungil;
 use pyo3::prelude::*;
 use std::{
    future::Future,
-    pin::Pin,
+    pin::{Pin, pin},
    task::{Context, Poll},
 };

@@ -25,13 +26,15 @@ where

 impl<F> Future for AllowThreads<F>
 where
-    F: Future + Send,
-    F::Output: Send,
+    F: Future + Ungil,
+    F::Output: Ungil,
 {
    type Output = F::Output;

    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
        let waker = cx.waker();
-        Python::attach(|py| py.detach(|| self.project().0.poll(&mut Context::from_waker(waker))))
+        Python::with_gil(|py| {
+            py.allow_threads(|| self.project().0.poll(&mut Context::from_waker(waker)))
+        })
    }
 }
@@ -0,0 +1,240 @@
+//! This module exists to hold examples of some pyo3 patterns that may be too complex to
+//! re-create from scratch, but too inhomogenous to create an abstraction/wrapper around.
+//!
+//! Pattern examples include:
+//!  - Async task handles: with GC-integrated cleanup
+//!  - Sync/async callbacks from python: with propper eventloop handling
+//!
+//! Mutability pattern: https://pyo3.rs/v0.26.0/async-await.html#send--static-constraint
+//!  - Store mutable fields in tokio's `Mutex<T>`
+//!  - For async code: take `&self` and `.lock().await`
+//!  - For sync code: take `&mut self` and `.get_mut()`
+
+use crate::ext::{PyResultExt as _, ResultExt as _, TokioRuntimeExt as _};
+use futures::FutureExt as _;
+use futures::future::BoxFuture;
+use pyo3::exceptions::PyRuntimeError;
+use pyo3::prelude::{PyModule, PyModuleMethods as _};
+use pyo3::{
+    Bound, Py, PyAny, PyErr, PyResult, PyTraverseError, PyVisit, Python, pyclass, pymethods,
+};
+use std::time::Duration;
+use tokio::sync::mpsc;
+use tokio::sync::mpsc::error::TryRecvError;
+
+fn needs_tokio_runtime() {
+    tokio::runtime::Handle::current();
+}
+
+type SyncCallback = Box<dyn Fn() + Send + Sync>;
+type AsyncCallback = Box<dyn Fn() -> BoxFuture<'static, ()> + Send + Sync>;
+
+enum AsyncTaskMessage {
+    SyncCallback(SyncCallback),
+    AsyncCallback(AsyncCallback),
+}
+
+async fn async_task(
+    sender: mpsc::UnboundedSender<()>,
+    mut receiver: mpsc::UnboundedReceiver<AsyncTaskMessage>,
+) {
+    log::info!("RUST: async task started");
+
+    // task state
+    let mut interval = tokio::time::interval(Duration::from_secs(1));
+
+    let mut sync_cbs: Vec<SyncCallback> = vec![];
+    let mut async_cbs: Vec<AsyncCallback> = vec![];
+
+    loop {
+        tokio::select! {
+            // handle incoming messages from task-handle
+            message = receiver.recv() => {
+                // handle closed channel by exiting
+                let Some(message) = message else {
+                    log::info!("RUST: channel closed");
+                    break;
+                };
+
+                // dispatch incoming event
+                match message {
+                    AsyncTaskMessage::SyncCallback(cb) => {
+                        sync_cbs.push(cb);
+                    }
+                    AsyncTaskMessage::AsyncCallback(cb) => {
+                        async_cbs.push(cb);
+                    }
+                }
+            }
+
+            // handle all other events
+            _ = interval.tick() => {
+                log::info!("RUST: async task tick");
+
+                // call back all sync callbacks
+                for cb in &sync_cbs {
+                    cb();
+                }
+
+                // call back all async callbacks
+                for cb in &async_cbs {
+                    cb().await;
+                }
+
+                // send event on unbounded channel
+                sender.send(()).expect("handle receiver cannot be closed/dropped");
+            }
+        }
+    }
+
+    log::info!("RUST: async task stopped");
+}
+
+// #[gen_stub_pyclass]
+#[pyclass(name = "AsyncTaskHandle")]
+#[derive(Debug)]
+struct PyAsyncTaskHandle {
+    sender: Option<mpsc::UnboundedSender<AsyncTaskMessage>>,
+    receiver: mpsc::UnboundedReceiver<()>,
+}
+
+#[allow(clippy::expect_used)]
+impl PyAsyncTaskHandle {
+    const fn sender(&self) -> &mpsc::UnboundedSender<AsyncTaskMessage> {
+        self.sender
+            .as_ref()
+            .expect("The sender should only be None after de-initialization.")
+    }
+
+    const fn sender_mut(&mut self) -> &mpsc::UnboundedSender<AsyncTaskMessage> {
+        self.sender
+            .as_mut()
+            .expect("The sender should only be None after de-initialization.")
+    }
+
+    const fn new(
+        sender: mpsc::UnboundedSender<AsyncTaskMessage>,
+        receiver: mpsc::UnboundedReceiver<()>,
+    ) -> Self {
+        Self {
+            sender: Some(sender),
+            receiver,
+        }
+    }
+}
+
+// #[gen_stub_pymethods]
+#[pymethods]
+impl PyAsyncTaskHandle {
+    #[new]
+    fn py_new(py: Python<'_>) -> PyResult<Self> {
+        use pyo3_async_runtimes::tokio::get_runtime;
+
+        // create communication channel TOWARDS our task
+        let (h_sender, t_receiver) = mpsc::unbounded_channel::<AsyncTaskMessage>();
+
+        // create communication channel FROM our task
+        let (t_sender, h_receiver) = mpsc::unbounded_channel::<()>();
+
+        // perform necessary setup within tokio context - or it crashes
+        let () = get_runtime().block_on(async { needs_tokio_runtime() });
+
+        // spawn tokio task with this thread's task-locals - without this, async callbacks on the new threads will not work!!
+        _ = get_runtime().spawn_with_scope(py, async move {
+            async_task(t_sender, t_receiver).await;
+        });
+        Ok(Self::new(h_sender, h_receiver))
+    }
+
+    /// NOTE: exceptions in callbacks are silently ignored until end of execution
+    fn add_sync_callback(
+        &self,
+        // #[gen_stub(override_type(
+        //     type_repr="collections.abc.Callable[[], None]",
+        //     imports=("collections.abc")
+        // ))]
+        callback: Py<PyAny>,
+    ) -> PyResult<()> {
+        // blocking call to async method -> can do non-blocking if needed
+        self.sender()
+            .send(AsyncTaskMessage::SyncCallback(Box::new(move || {
+                _ = Python::with_gil(|py| callback.call0(py).write_unraisable_with(py));
+            })))
+            .pyerr()?;
+        Ok(())
+    }
+
+    /// NOTE: exceptions in callbacks are silently ignored until end of execution
+    fn add_async_callback(
+        &self,
+        // #[gen_stub(override_type(
+        //     type_repr="collections.abc.Callable[[], collections.abc.Awaitable[None]]",
+        //     imports=("collections.abc")
+        // ))]
+        callback: Py<PyAny>,
+    ) -> PyResult<()> {
+        // blocking call to async method -> can do non-blocking if needed
+        self.sender()
+            .send(AsyncTaskMessage::AsyncCallback(Box::new(move || {
+                let c = Python::with_gil(|py| callback.clone_ref(py));
+                async move {
+                    if let Some(f) = Python::with_gil(|py| {
+                        let coroutine = c.call0(py).write_unraisable_with(py)?;
+                        pyo3_async_runtimes::tokio::into_future(coroutine.into_bound(py))
+                            .write_unraisable_with(py)
+                    }) {
+                        _ = f.await.write_unraisable();
+                    }
+                }
+                .boxed()
+            })))
+            .pyerr()?;
+        Ok(())
+    }
+
+    async fn receive_unit(&mut self) -> PyResult<()> {
+        self.receiver
+            .recv()
+            .await
+            .ok_or(PyErr::new::<PyRuntimeError, _>(
+                "cannot receive unit on closed channel",
+            ))
+    }
+
+    fn drain_units(&mut self) -> PyResult<i32> {
+        let mut cnt = 0;
+        loop {
+            match self.receiver.try_recv() {
+                Err(TryRecvError::Disconnected) => {
+                    return Err(PyErr::new::<PyRuntimeError, _>(
+                        "cannot receive unit on closed channel",
+                    ));
+                }
+                Err(TryRecvError::Empty) => return Ok(cnt),
+                Ok(()) => {
+                    cnt += 1;
+                    continue;
+                }
+            }
+        }
+    }
+
+    // #[gen_stub(skip)]
+    const fn __traverse__(&self, _visit: PyVisit<'_>) -> Result<(), PyTraverseError> {
+        Ok(()) // This is needed purely so `__clear__` can work
+    }
+
+    // #[gen_stub(skip)]
+    fn __clear__(&mut self) {
+        // TODO: may or may not need to await a "kill-signal" oneshot channel message,
+        //       to ensure that the networking task is done BEFORE exiting the clear function...
+        //       but this may require GIL?? and it may not be safe to call GIL here??
+        self.sender = None; // Using Option<T> as a trick to force `sender` channel to be dropped
+    }
+}
+
+pub fn examples_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
+    m.add_class::<PyAsyncTaskHandle>()?;
+
+    Ok(())
+}
@@ -1,47 +0,0 @@
-use crate::ext::ResultExt as _;
-use libp2p::identity::Keypair;
-use pyo3::types::{PyBytes, PyBytesMethods as _};
-use pyo3::{Bound, PyResult, Python, pyclass, pymethods};
-use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
-
-/// Identity keypair of a node.
-#[gen_stub_pyclass]
-#[pyclass(name = "Keypair", frozen)]
-#[repr(transparent)]
-pub struct PyKeypair(pub Keypair);
-
-#[gen_stub_pymethods]
-#[pymethods]
-#[allow(clippy::needless_pass_by_value)]
-impl PyKeypair {
-    /// Generate a new Ed25519 keypair.
-    #[staticmethod]
-    fn generate() -> Self {
-        Self(Keypair::generate_ed25519())
-    }
-
-    /// Construct an Ed25519 keypair from secret key bytes
-    #[staticmethod]
-    fn from_bytes(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
-        let mut bytes = Vec::from(bytes.as_bytes());
-        Ok(Self(Keypair::ed25519_from_bytes(&mut bytes).pyerr()?))
-    }
-
-    /// Get the secret key bytes underlying the keypair
-    fn to_bytes<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyBytes>> {
-        let bytes = self
-            .0
-            .clone()
-            .try_into_ed25519()
-            .pyerr()?
-            .secret()
-            .as_ref()
-            .to_vec();
-        Ok(PyBytes::new(py, &bytes))
-    }
-
-    /// Convert the `Keypair` into the corresponding `PeerId` string, which we use as our `NodeId`.
-    fn to_node_id(&self) -> String {
-        self.0.public().to_peer_id().to_base58()
-    }
-}
@@ -4,14 +4,28 @@
 //!
 //!

-mod allow_threading;
-mod ident;
-mod networking;
+// enable Rust-unstable features for convenience
+#![feature(trait_alias)]
+#![feature(tuple_trait)]
+#![feature(unboxed_closures)]
+// #![feature(stmt_expr_attributes)]
+// #![feature(assert_matches)]
+// #![feature(async_fn_in_dyn_trait)]
+// #![feature(async_for_loop)]
+// #![feature(auto_traits)]
+// #![feature(negative_impls)]
+
+extern crate core;
+mod allow_threading;
+mod examples;
+pub(crate) mod networking;
+pub(crate) mod pylibp2p;

-use crate::ident::PyKeypair;
 use crate::networking::networking_submodule;
+use crate::pylibp2p::ident::ident_submodule;
+use crate::pylibp2p::multiaddr::multiaddr_submodule;
 use pyo3::prelude::PyModule;
-use pyo3::types::PyModuleMethods;
+use pyo3::prelude::*;
 use pyo3::{Bound, PyResult, pyclass, pymodule};
 use pyo3_stub_gen::define_stub_info_gatherer;

@@ -20,11 +34,24 @@ pub(crate) mod r#const {
    pub const MPSC_CHANNEL_SIZE: usize = 1024;
 }

+/// Namespace for all the type/trait aliases used by this crate.
+pub(crate) mod alias {
+    use std::error::Error;
+    use std::marker::Tuple;
+
+    pub trait SendFn<Args: Tuple + Send + 'static, Output> =
+        Fn<Args, Output = Output> + Send + 'static;
+
+    pub type AnyError = Box<dyn Error + Send + Sync + 'static>;
+    pub type AnyResult<T> = Result<T, AnyError>;
+}
+
 /// Namespace for crate-wide extension traits/methods
 pub(crate) mod ext {
    use crate::allow_threading::AllowThreads;
    use extend::ext;
    use pyo3::exceptions::{PyConnectionError, PyRuntimeError};
+    use pyo3::marker::Ungil;
    use pyo3::types::PyBytes;
    use pyo3::{Py, PyErr, PyResult, Python};
    use tokio::runtime::Runtime;
@@ -35,7 +62,7 @@ pub(crate) mod ext {
    #[ext(pub, name = ByteArrayExt)]
    impl [u8] {
        fn pybytes(&self) -> Py<PyBytes> {
-            Python::attach(|py| PyBytes::new(py, self).unbind())
+            Python::with_gil(|py| PyBytes::new(py, self).unbind())
        }
    }

@@ -71,7 +98,7 @@ pub(crate) mod ext {
    #[ext(pub, name = PyResultExt)]
    impl<T> PyResult<T> {
        fn write_unraisable(self) -> Option<T> {
-            Python::attach(|py| self.write_unraisable_with(py))
+            Python::with_gil(|py| self.write_unraisable_with(py))
        }

        fn write_unraisable_with(self, py: Python<'_>) -> Option<T> {
@@ -148,6 +175,24 @@ pub(crate) mod ext {
    }
 }

+pub(crate) mod private {
+    use std::marker::Sized;
+
+    /// Sealed traits support
+    pub trait Sealed {}
+    impl<T: ?Sized> Sealed for T {}
+}
+
+/// A wrapper around [`Py`] that implements [`Clone`] using [`Python::with_gil`].
+#[repr(transparent)]
+pub(crate) struct ClonePy<T>(pub Py<T>);
+
+impl<T> Clone for ClonePy<T> {
+    fn clone(&self) -> Self {
+        Python::with_gil(|py| Self(self.0.clone_ref(py)))
+    }
+}
+
 /// A Python module implemented in Rust. The name of this function must match
 /// the `lib.name` setting in the `Cargo.toml`, else Python will not be able to
 /// import the module.
@@ -159,7 +204,8 @@ fn main_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    // TODO: for now this is all NOT a submodule, but figure out how to make the submodule system
    //       work with maturin, where the types generate correctly, in the right folder, without
    //       too many importing issues...
-    m.add_class::<PyKeypair>()?;
+    ident_submodule(m)?;
+    multiaddr_submodule(m)?;
    networking_submodule(m)?;

    // top-level constructs
@@ -8,12 +8,12 @@
 use crate::r#const::MPSC_CHANNEL_SIZE;
 use crate::ext::{ByteArrayExt as _, FutureExt, PyErrExt as _};
 use crate::ext::{ResultExt as _, TokioMpscReceiverExt as _, TokioMpscSenderExt as _};
-use crate::ident::PyKeypair;
 use crate::pyclass;
+use crate::pylibp2p::ident::{PyKeypair, PyPeerId};
 use libp2p::futures::StreamExt as _;
-use libp2p::gossipsub;
 use libp2p::gossipsub::{IdentTopic, Message, MessageId, PublishError};
 use libp2p::swarm::SwarmEvent;
+use libp2p::{gossipsub, mdns};
 use networking::discovery;
 use networking::swarm::create_swarm;
 use pyo3::prelude::{PyModule, PyModuleMethods as _};
@@ -25,7 +25,7 @@ use tokio::sync::{Mutex, mpsc, oneshot};

 mod exception {
    use pyo3::types::PyTuple;
-    use pyo3::{exceptions::PyException, prelude::*};
+    use pyo3::{PyErrArguments, exceptions::PyException, prelude::*};
    use pyo3_stub_gen::derive::*;

    #[gen_stub_pyclass]
@@ -119,7 +119,7 @@ struct PyConnectionUpdate {

    /// Identity of the peer that we have connected to or disconnected from.
    #[pyo3(get)]
-    peer_id: String,
+    peer_id: PyPeerId,

    /// Remote connection's IPv4 address.
    #[pyo3(get)]
@@ -155,6 +155,7 @@ async fn networking_task(
 ) {
    use SwarmEvent::*;
    use ToTask::*;
+    use mdns::Event::*;
    use networking::swarm::BehaviourEvent::*;

    log::info!("RUST: networking task started");
@@ -251,7 +252,7 @@ async fn networking_task(
                        // send connection event to channel (or exit if connection closed)
                        if let Err(e) = connection_update_tx.send(PyConnectionUpdate {
                            update_type: PyConnectionUpdateType::Connected,
-                            peer_id: peer_id.to_base58(),
+                            peer_id: PyPeerId(peer_id),
                            remote_ipv4,
                            remote_tcp_port,
                        }).await {
@@ -272,7 +273,7 @@ async fn networking_task(
                        // send disconnection event to channel (or exit if connection closed)
                        if let Err(e) = connection_update_tx.send(PyConnectionUpdate {
                            update_type: PyConnectionUpdateType::Disconnected,
-                            peer_id: peer_id.to_base58(),
+                            peer_id: PyPeerId(peer_id),
                            remote_ipv4,
                            remote_tcp_port,
                        }).await {
@@ -341,8 +342,10 @@ impl PyNetworkingHandle {
    // ---- Lifecycle management methods ----

    #[new]
-    fn py_new(identity: Bound<'_, PyKeypair>) -> PyResult<Self> {
+    #[pyo3(signature = (identity, bootstrap_peers=vec![]))]
+    fn py_new(identity: Bound<'_, PyKeypair>, bootstrap_peers: Vec<String>) -> PyResult<Self> {
        use pyo3_async_runtimes::tokio::get_runtime;
+        use std::str::FromStr;

        // create communication channels
        let (to_task_tx, to_task_rx) = mpsc::channel(MPSC_CHANNEL_SIZE);
@@ -352,9 +355,20 @@ impl PyNetworkingHandle {
        // get identity
        let identity = identity.borrow().0.clone();

+        // parse bootstrap peer multiaddrs
+        let parsed_peers: Vec<libp2p::Multiaddr> = bootstrap_peers
+            .into_iter()
+            .map(|s| libp2p::Multiaddr::from_str(&s))
+            .collect::<Result<_, _>>()
+            .pyerr()?;
+
+        if !parsed_peers.is_empty() {
+            log::info!("RUST: bootstrap peers: {:?}", parsed_peers);
+        }
+
        // create networking swarm (within tokio context!! or it crashes)
        let swarm = get_runtime()
-            .block_on(async { create_swarm(identity) })
+            .block_on(async { create_swarm(identity, parsed_peers) })
            .pyerr()?;

        // spawn tokio task running the networking logic
@@ -484,7 +498,7 @@ impl PyNetworkingHandle {
        let (tx, rx) = oneshot::channel();

        // send off request to subscribe
-        let data = Python::attach(|py| Vec::from(data.as_bytes(py)));
+        let data = Python::with_gil(|py| Vec::from(data.as_bytes(py)));
        self.to_task_tx()
            .send_py(ToTask::GossipsubPublish {
                topic,
@@ -0,0 +1,159 @@
+use crate::ext::ResultExt as _;
+use libp2p::PeerId;
+use libp2p::identity::Keypair;
+use pyo3::prelude::{PyBytesMethods as _, PyModule, PyModuleMethods as _};
+use pyo3::types::PyBytes;
+use pyo3::{Bound, PyResult, Python, pyclass, pymethods};
+use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
+
+/// Identity keypair of a node.
+#[gen_stub_pyclass]
+#[pyclass(name = "Keypair", frozen)]
+#[repr(transparent)]
+pub struct PyKeypair(pub Keypair);
+
+#[gen_stub_pymethods]
+#[pymethods]
+#[allow(clippy::needless_pass_by_value)]
+impl PyKeypair {
+    /// Generate a new Ed25519 keypair.
+    #[staticmethod]
+    fn generate_ed25519() -> Self {
+        Self(Keypair::generate_ed25519())
+    }
+
+    /// Generate a new ECDSA keypair.
+    #[staticmethod]
+    fn generate_ecdsa() -> Self {
+        Self(Keypair::generate_ecdsa())
+    }
+
+    /// Generate a new Secp256k1 keypair.
+    #[staticmethod]
+    fn generate_secp256k1() -> Self {
+        Self(Keypair::generate_secp256k1())
+    }
+
+    /// Decode a private key from a protobuf structure and parse it as a `Keypair`.
+    #[staticmethod]
+    fn from_protobuf_encoding(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(Keypair::from_protobuf_encoding(&bytes).pyerr()?))
+    }
+
+    /// Decode an keypair from a DER-encoded secret key in PKCS#8 `PrivateKeyInfo`
+    /// format (i.e. unencrypted) as defined in [RFC5208].
+    ///
+    /// [RFC5208]: https://tools.ietf.org/html/rfc5208#section-5
+    #[staticmethod]
+    fn rsa_from_pkcs8(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let mut bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(Keypair::rsa_from_pkcs8(&mut bytes).pyerr()?))
+    }
+
+    /// Decode a keypair from a DER-encoded Secp256k1 secret key in an `ECPrivateKey`
+    /// structure as defined in [RFC5915].
+    ///
+    /// [RFC5915]: https://tools.ietf.org/html/rfc5915
+    #[staticmethod]
+    fn secp256k1_from_der(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let mut bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(Keypair::secp256k1_from_der(&mut bytes).pyerr()?))
+    }
+
+    #[staticmethod]
+    fn ed25519_from_bytes(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let mut bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(Keypair::ed25519_from_bytes(&mut bytes).pyerr()?))
+    }
+
+    /// Encode a private key as protobuf structure.
+    fn to_protobuf_encoding<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyBytes>> {
+        let bytes = self.0.to_protobuf_encoding().pyerr()?;
+        Ok(PyBytes::new(py, &bytes))
+    }
+
+    /// Convert the `Keypair` into the corresponding `PeerId`.
+    fn to_peer_id(&self) -> PyPeerId {
+        PyPeerId(self.0.public().to_peer_id())
+    }
+
+    // /// Hidden constructor for pickling support. TODO: figure out how to do pickling...
+    // #[gen_stub(skip)]
+    // #[new]
+    // fn py_new(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+    //     Self::from_protobuf_encoding(bytes)
+    // }
+    //
+    // #[gen_stub(skip)]
+    // fn __setstate__(&mut self, state: Bound<'_, PyBytes>) -> PyResult<()> {
+    //     *self = Self::from_protobuf_encoding(state)?;
+    //     Ok(())
+    // }
+    //
+    // #[gen_stub(skip)]
+    // fn __getstate__<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyBytes>> {
+    //     self.to_protobuf_encoding(py)
+    // }
+    //
+    // #[gen_stub(skip)]
+    // pub fn __getnewargs__<'py>(&self, py: Python<'py>) -> PyResult<(Bound<'py, PyBytes>,)> {
+    //     Ok((self.to_protobuf_encoding(py)?,))
+    // }
+}
+
+/// Identifier of a peer of the network.
+///
+/// The data is a `CIDv0` compatible multihash of the protobuf encoded public key of the peer
+/// as specified in [specs/peer-ids](https://github.com/libp2p/specs/blob/master/peer-ids/peer-ids.md).
+#[gen_stub_pyclass]
+#[pyclass(name = "PeerId", frozen)]
+#[derive(Debug, Clone)]
+#[repr(transparent)]
+pub struct PyPeerId(pub PeerId);
+
+#[gen_stub_pymethods]
+#[pymethods]
+#[allow(clippy::needless_pass_by_value)]
+impl PyPeerId {
+    /// Generates a random peer ID from a cryptographically secure PRNG.
+    ///
+    /// This is useful for randomly walking on a DHT, or for testing purposes.
+    #[staticmethod]
+    fn random() -> Self {
+        Self(PeerId::random())
+    }
+
+    /// Parses a `PeerId` from bytes.
+    #[staticmethod]
+    fn from_bytes(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(PeerId::from_bytes(&bytes).pyerr()?))
+    }
+
+    /// Returns a raw bytes representation of this `PeerId`.
+    fn to_bytes<'py>(&self, py: Python<'py>) -> Bound<'py, PyBytes> {
+        let bytes = self.0.to_bytes();
+        PyBytes::new(py, &bytes)
+    }
+
+    /// Returns a base-58 encoded string of this `PeerId`.
+    fn to_base58(&self) -> String {
+        self.0.to_base58()
+    }
+
+    fn __repr__(&self) -> String {
+        format!("PeerId({})", self.to_base58())
+    }
+
+    fn __str__(&self) -> String {
+        self.to_base58()
+    }
+}
+
+pub fn ident_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
+    m.add_class::<PyKeypair>()?;
+    m.add_class::<PyPeerId>()?;
+
+    Ok(())
+}
@@ -0,0 +1,8 @@
+//! A module for exposing Rust's libp2p datatypes over Pyo3
+//!
+//! TODO: right now we are coupled to libp2p's identity, but eventually we want to create our own
+//!       independent identity type of some kind or another. This may require handshaking.
+//!
+
+pub mod ident;
+pub mod multiaddr;
@@ -0,0 +1,81 @@
+use crate::ext::ResultExt as _;
+use libp2p::Multiaddr;
+use pyo3::prelude::{PyBytesMethods as _, PyModule, PyModuleMethods as _};
+use pyo3::types::PyBytes;
+use pyo3::{Bound, PyResult, Python, pyclass, pymethods};
+use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
+use std::str::FromStr as _;
+
+/// Representation of a Multiaddr.
+#[gen_stub_pyclass]
+#[pyclass(name = "Multiaddr", frozen)]
+#[derive(Debug, Clone)]
+#[repr(transparent)]
+pub struct PyMultiaddr(pub Multiaddr);
+
+#[gen_stub_pymethods]
+#[pymethods]
+#[allow(clippy::needless_pass_by_value)]
+impl PyMultiaddr {
+    /// Create a new, empty multiaddress.
+    #[staticmethod]
+    fn empty() -> Self {
+        Self(Multiaddr::empty())
+    }
+
+    /// Create a new, empty multiaddress with the given capacity.
+    #[staticmethod]
+    fn with_capacity(n: usize) -> Self {
+        Self(Multiaddr::with_capacity(n))
+    }
+
+    /// Parse a `Multiaddr` value from its byte slice representation.
+    #[staticmethod]
+    fn from_bytes(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(Multiaddr::try_from(bytes).pyerr()?))
+    }
+
+    /// Parse a `Multiaddr` value from its string representation.
+    #[staticmethod]
+    fn from_string(string: String) -> PyResult<Self> {
+        Ok(Self(Multiaddr::from_str(&string).pyerr()?))
+    }
+
+    /// Return the length in bytes of this multiaddress.
+    fn len(&self) -> usize {
+        self.0.len()
+    }
+
+    /// Returns true if the length of this multiaddress is 0.
+    fn is_empty(&self) -> bool {
+        self.0.is_empty()
+    }
+
+    /// Return a copy of this [`Multiaddr`]'s byte representation.
+    fn to_bytes<'py>(&self, py: Python<'py>) -> Bound<'py, PyBytes> {
+        let bytes = self.0.to_vec();
+        PyBytes::new(py, &bytes)
+    }
+
+    /// Convert a Multiaddr to a string.
+    fn to_string(&self) -> String {
+        self.0.to_string()
+    }
+
+    #[gen_stub(skip)]
+    fn __repr__(&self) -> String {
+        format!("Multiaddr({})", self.0)
+    }
+
+    #[gen_stub(skip)]
+    fn __str__(&self) -> String {
+        self.to_string()
+    }
+}
+
+pub fn multiaddr_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
+    m.add_class::<PyMultiaddr>()?;
+
+    Ok(())
+}
@@ -19,14 +19,21 @@ either = { workspace = true }
 # macro dependencies
 extend = { workspace = true }
 delegate = { workspace = true }
+impl-trait-for-tuples = { workspace = true }
+derive_more = { workspace = true }

 # async
 tokio = { workspace = true, features = ["full"] }
-futures-lite = { workspace = true }
+futures = { workspace = true }
 futures-timer = { workspace = true }

 # utility dependencies
 util = { workspace = true }
+thiserror = { workspace = true }
+#internment = { workspace = true }
+#recursion = { workspace = true }
+#generativity = { workspace = true }
+#itertools = { workspace = true }
 tracing-subscriber = { version = "0.3.19", features = ["default", "env-filter"] }
 keccak-const = { workspace = true }

@@ -34,4 +41,4 @@ keccak-const = { workspace = true }
 log = { workspace = true }

 # networking
-libp2p = { workspace = true, features = ["full"] }
+libp2p = { workspace = true, features = ["full"] }
@@ -1,4 +1,4 @@
-use futures_lite::StreamExt;
+use futures::stream::StreamExt as _;
 use libp2p::{gossipsub, identity, swarm::SwarmEvent};
 use networking::{discovery, swarm};
 use tokio::{io, io::AsyncBufReadExt as _, select};
@@ -38,19 +38,19 @@ async fn main() {
                    println!("Publish error: {e:?}");
                }
            }
-            event = swarm.next() => match event {
+            event = swarm.select_next_some() => match event {
                // on gossipsub incoming
-                Some(SwarmEvent::Behaviour(swarm::BehaviourEvent::Gossipsub(gossipsub::Event::Message {
+                SwarmEvent::Behaviour(swarm::BehaviourEvent::Gossipsub(gossipsub::Event::Message {
                    propagation_source: peer_id,
                    message_id: id,
                    message,
-                }))) => println!(
+                })) => println!(
                        "\n\nGot message: '{}' with id: {id} from peer: {peer_id}\n\n",
                        String::from_utf8_lossy(&message.data),
                    ),

                // on discovery
-                Some(SwarmEvent::Behaviour(swarm::BehaviourEvent::Discovery(e)) )=> match e {
+                SwarmEvent::Behaviour(swarm::BehaviourEvent::Discovery(e)) => match e {
                    discovery::Event::ConnectionEstablished {
                        peer_id, connection_id, remote_ip, remote_tcp_port
                    } => {
@@ -64,7 +64,7 @@ async fn main() {
                }

                // ignore outgoing errors: those are normal
-                e@Some(SwarmEvent::OutgoingConnectionError { .. }) => { log::debug!("Outgoing connection error: {e:?}"); }
+                e@SwarmEvent::OutgoingConnectionError { .. } => { log::debug!("Outgoing connection error: {e:?}"); }

                // otherwise log any other event
                e => { log::info!("Other event {e:?}"); }
@@ -0,0 +1,127 @@
+// Copyright 2018 Parity Technologies (UK) Ltd.
+//
+// Permission is hereby granted, free of charge, to any person obtaining a
+// copy of this software and associated documentation files (the "Software"),
+// to deal in the Software without restriction, including without limitation
+// the rights to use, copy, modify, merge, publish, distribute, sublicense,
+// and/or sell copies of the Software, and to permit persons to whom the
+// Software is furnished to do so, subject to the following conditions:
+//
+// The above copyright notice and this permission notice shall be included in
+// all copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+// OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+// DEALINGS IN THE SOFTWARE.
+
+use futures::stream::StreamExt;
+use libp2p::{
+    gossipsub, mdns, noise,
+    swarm::{NetworkBehaviour, SwarmEvent},
+    tcp, yamux,
+};
+use std::time::Duration;
+use std::{error::Error, hash::Hash};
+use tokio::{io, io::AsyncBufReadExt, select};
+use tracing_subscriber::EnvFilter;
+
+// We create a custom network behaviour that combines Gossipsub and Mdns.
+#[derive(NetworkBehaviour)]
+struct MyBehaviour {
+    gossipsub: gossipsub::Behaviour,
+    mdns: mdns::tokio::Behaviour,
+}
+
+#[tokio::main]
+async fn main() -> Result<(), Box<dyn Error>> {
+    let _ = tracing_subscriber::fmt()
+        .with_env_filter(EnvFilter::from_default_env())
+        .try_init();
+
+    let mut swarm = libp2p::SwarmBuilder::with_new_identity()
+        .with_tokio()
+        .with_tcp(
+            tcp::Config::default(),
+            noise::Config::new,
+            yamux::Config::default,
+        )?
+        .with_behaviour(|key| {
+            // Set a custom gossipsub configuration
+            let gossipsub_config = gossipsub::ConfigBuilder::default()
+                .heartbeat_interval(Duration::from_secs(10))
+                .validation_mode(gossipsub::ValidationMode::Strict) // This sets the kind of message validation. The default is Strict (enforce message signing)
+                .build()
+                .map_err(io::Error::other)?; // Temporary hack because `build` does not return a proper `std::error::Error`.
+
+            // build a gossipsub network behaviour
+            let gossipsub = gossipsub::Behaviour::new(
+                gossipsub::MessageAuthenticity::Signed(key.clone()),
+                gossipsub_config,
+            )?;
+
+            let mdns =
+                mdns::tokio::Behaviour::new(mdns::Config::default(), key.public().to_peer_id())?;
+            Ok(MyBehaviour { gossipsub, mdns })
+        })?
+        .build();
+
+    println!("Running swarm with identity {}", swarm.local_peer_id());
+
+    // Create a Gossipsub topic
+    let topic = gossipsub::IdentTopic::new("test-net");
+    // subscribes to our topic
+    swarm.behaviour_mut().gossipsub.subscribe(&topic)?;
+
+    // Read full lines from stdin
+    let mut stdin = io::BufReader::new(io::stdin()).lines();
+
+    // Listen on all interfaces and whatever port the OS assigns
+    swarm.listen_on("/ip4/0.0.0.0/tcp/0".parse()?)?;
+
+    println!("Enter messages via STDIN and they will be sent to connected peers using Gossipsub");
+
+    // Kick it off
+    loop {
+        select! {
+            Ok(Some(line)) = stdin.next_line() => {
+                if let Err(e) = swarm
+                    .behaviour_mut().gossipsub
+                    .publish(topic.clone(), line.as_bytes()) {
+                    println!("Publish error: {e:?}");
+                }
+            }
+            event = swarm.select_next_some() => match event {
+                SwarmEvent::Behaviour(MyBehaviourEvent::Mdns(mdns::Event::Discovered(list))) => {
+                    for (peer_id, multiaddr) in list {
+                        println!("mDNS discovered a new peer: {peer_id} on {multiaddr}");
+                        swarm.behaviour_mut().gossipsub.add_explicit_peer(&peer_id);
+                    }
+                },
+                SwarmEvent::Behaviour(MyBehaviourEvent::Mdns(mdns::Event::Expired(list))) => {
+                    for (peer_id, multiaddr) in list {
+                        println!("mDNS discover peer has expired: {peer_id} on {multiaddr}");
+                        swarm.behaviour_mut().gossipsub.remove_explicit_peer(&peer_id);
+                    }
+                },
+                SwarmEvent::Behaviour(MyBehaviourEvent::Gossipsub(gossipsub::Event::Message {
+                    propagation_source: peer_id,
+                    message_id: id,
+                    message,
+                })) => println!(
+                        "Got message: '{}' with id: {id} from peer: {peer_id}",
+                        String::from_utf8_lossy(&message.data),
+                    ),
+                SwarmEvent::NewListenAddr { address, .. } => {
+                    println!("Local node is listening on {address}");
+                }
+                e => {
+                    println!("Other swarm event: {:?}", e);
+                }
+            }
+        }
+    }
+}
@@ -1,7 +1,8 @@
 use crate::ext::MultiaddrExt;
+use crate::keep_alive;
 use delegate::delegate;
 use either::Either;
-use futures_lite::FutureExt;
+use futures::FutureExt;
 use futures_timer::Delay;
 use libp2p::core::transport::PortUse;
 use libp2p::core::{ConnectedPoint, Endpoint};
@@ -105,6 +106,9 @@ pub struct Behaviour {
    managed: managed::Behaviour,
    mdns_discovered: HashMap<PeerId, BTreeSet<Multiaddr>>,

+    /// Addresses provided via --bootstrap-peer, re-dialed on the retry loop.
+    bootstrap_addrs: Vec<Multiaddr>,
+
    retry_delay: Delay, // retry interval

    // pending events to emmit => waker-backed Deque to control polling
@@ -112,13 +116,21 @@ pub struct Behaviour {
 }

 impl Behaviour {
-    pub fn new(keypair: &identity::Keypair) -> io::Result<Self> {
-        Ok(Self {
+    pub fn new(keypair: &identity::Keypair, bootstrap_peers: Vec<Multiaddr>) -> io::Result<Self> {
+        let mut behaviour = Self {
            managed: managed::Behaviour::new(keypair)?,
            mdns_discovered: HashMap::new(),
+            bootstrap_addrs: bootstrap_peers,
            retry_delay: Delay::new(RETRY_CONNECT_INTERVAL),
            pending_events: WakerDeque::new(),
-        })
+        };
+
+        // Immediately dial all bootstrap peers
+        for addr in &behaviour.bootstrap_addrs.clone() {
+            behaviour.dial_addr(addr.clone());
+        }
+
+        Ok(behaviour)
    }

    fn dial(&mut self, peer_id: PeerId, addr: Multiaddr) {
@@ -127,6 +139,14 @@ impl Behaviour {
        })
    }

+    /// Dial by address only — PeerId is resolved via Noise handshake on connect.
+    /// Used for bootstrap peers where only IP:port is known.
+    fn dial_addr(&mut self, addr: Multiaddr) {
+        self.pending_events.push_back(ToSwarm::Dial {
+            opts: DialOpts::unknown_peer_id().address(addr).build(),
+        })
+    }
+
    fn close_connection(&mut self, peer_id: PeerId, connection: ConnectionId) {
        // push front to make this IMMEDIATE
        self.pending_events.push_front(ToSwarm::CloseConnection {
@@ -361,13 +381,16 @@ impl NetworkBehaviour for Behaviour {
            Poll::Pending => {}
        }

-        // retry connecting to all mDNS peers periodically (fails safely if already connected)
-        if self.retry_delay.poll(cx).is_ready() {
+        // retry connecting to all mDNS + bootstrap peers periodically (fails safely if already connected)
+        if self.retry_delay.poll_unpin(cx).is_ready() {
            for (p, mas) in self.mdns_discovered.clone() {
                for ma in mas {
                    self.dial(p, ma)
                }
            }
+            for addr in self.bootstrap_addrs.clone() {
+                self.dial_addr(addr)
+            }
            self.retry_delay.reset(RETRY_CONNECT_INTERVAL) // reset timeout
        }

@@ -0,0 +1,44 @@
+use delegate::delegate;
+use libp2p::swarm::handler::ConnectionEvent;
+use libp2p::swarm::{ConnectionHandlerEvent, SubstreamProtocol, dummy, handler};
+use std::task::{Context, Poll};
+
+/// An implementation of [`ConnectionHandler`] that doesn't handle any protocols, but it keeps
+/// the connection alive.
+#[derive(Clone)]
+#[repr(transparent)]
+pub struct ConnectionHandler(dummy::ConnectionHandler);
+
+impl ConnectionHandler {
+    pub fn new() -> Self {
+        ConnectionHandler(dummy::ConnectionHandler)
+    }
+}
+
+impl handler::ConnectionHandler for ConnectionHandler {
+    // delegate types and implementation mostly to dummy handler
+    type FromBehaviour = <dummy::ConnectionHandler as handler::ConnectionHandler>::FromBehaviour;
+    type ToBehaviour = <dummy::ConnectionHandler as handler::ConnectionHandler>::ToBehaviour;
+    type InboundProtocol =
+        <dummy::ConnectionHandler as handler::ConnectionHandler>::InboundProtocol;
+    type OutboundProtocol =
+        <dummy::ConnectionHandler as handler::ConnectionHandler>::OutboundProtocol;
+    type InboundOpenInfo =
+        <dummy::ConnectionHandler as handler::ConnectionHandler>::InboundOpenInfo;
+    type OutboundOpenInfo =
+        <dummy::ConnectionHandler as handler::ConnectionHandler>::OutboundOpenInfo;
+
+    delegate! {
+        to self.0 {
+            fn listen_protocol(&self) -> SubstreamProtocol<Self::InboundProtocol, Self::InboundOpenInfo>;
+            fn poll(&mut self, cx: &mut Context<'_>) -> Poll<ConnectionHandlerEvent<Self::OutboundProtocol, Self::OutboundOpenInfo, Self::ToBehaviour>>;
+            fn on_behaviour_event(&mut self, event: Self::FromBehaviour);
+            fn on_connection_event(&mut self, event: ConnectionEvent<Self::InboundProtocol, Self::OutboundProtocol, Self::InboundOpenInfo, Self::OutboundOpenInfo>);
+        }
+    }
+
+    // specifically override this to force connection to stay alive
+    fn connection_keep_alive(&self) -> bool {
+        true
+    }
+}
@@ -3,7 +3,19 @@
 //! this is here as a placeholder documentation
 //!
 //!
+
+// enable Rust-unstable features for convenience
+#![feature(trait_alias)]
+// #![feature(stmt_expr_attributes)]
+// #![feature(unboxed_closures)]
+// #![feature(assert_matches)]
+// #![feature(async_fn_in_dyn_trait)]
+// #![feature(async_for_loop)]
+// #![feature(auto_traits)]
+// #![feature(negative_impls)]
+
 pub mod discovery;
+pub mod keep_alive;
 pub mod swarm;

 /// Namespace for all the type/trait aliases used by this crate.
@@ -42,3 +54,11 @@ pub(crate) mod ext {
        }
    }
 }
+
+pub(crate) mod private {
+    #![allow(dead_code)]
+
+    /// Sealed traits support
+    pub trait Sealed {}
+    impl<T: ?Sized> Sealed for T {}
+}
@@ -1,7 +1,7 @@
 use crate::alias;
 use crate::swarm::transport::tcp_transport;
 pub use behaviour::{Behaviour, BehaviourEvent};
-use libp2p::{SwarmBuilder, identity};
+use libp2p::{Multiaddr, SwarmBuilder, identity};

 pub type Swarm = libp2p::Swarm<Behaviour>;

@@ -16,11 +16,14 @@ pub const NETWORK_VERSION: &[u8] = b"v0.0.1";
 pub const OVERRIDE_VERSION_ENV_VAR: &str = "EXO_LIBP2P_NAMESPACE";

 /// Create and configure a swarm which listens to all ports on OS
-pub fn create_swarm(keypair: identity::Keypair) -> alias::AnyResult<Swarm> {
+pub fn create_swarm(
+    keypair: identity::Keypair,
+    bootstrap_peers: Vec<Multiaddr>,
+) -> alias::AnyResult<Swarm> {
    let mut swarm = SwarmBuilder::with_existing_identity(keypair)
        .with_tokio()
        .with_other_transport(tcp_transport)?
-        .with_behaviour(Behaviour::new)?
+        .with_behaviour(|kp| Behaviour::new(kp, bootstrap_peers))?
        .build();

    // Listen on all interfaces and whatever port the OS assigns
@@ -31,7 +34,7 @@ pub fn create_swarm(keypair: identity::Keypair) -> alias::AnyResult<Swarm> {
 mod transport {
    use crate::alias;
    use crate::swarm::{NETWORK_VERSION, OVERRIDE_VERSION_ENV_VAR};
-    use futures_lite::{AsyncRead, AsyncWrite};
+    use futures::{AsyncRead, AsyncWrite};
    use keccak_const::Sha3_256;
    use libp2p::core::muxing;
    use libp2p::core::transport::Boxed;
@@ -105,7 +108,7 @@ mod transport {
 mod behaviour {
    use crate::{alias, discovery};
    use libp2p::swarm::NetworkBehaviour;
-    use libp2p::{gossipsub, identity};
+    use libp2p::{Multiaddr, gossipsub, identity};

    /// Behavior of the Swarm which composes all desired behaviors:
    /// Right now its just [`discovery::Behaviour`] and [`gossipsub::Behaviour`].
@@ -116,9 +119,12 @@ mod behaviour {
    }

    impl Behaviour {
-        pub fn new(keypair: &identity::Keypair) -> alias::AnyResult<Self> {
+        pub fn new(
+            keypair: &identity::Keypair,
+            bootstrap_peers: Vec<Multiaddr>,
+        ) -> alias::AnyResult<Self> {
            Ok(Self {
-                discovery: discovery::Behaviour::new(keypair)?,
+                discovery: discovery::Behaviour::new(keypair, bootstrap_peers)?,
                gossipsub: gossipsub_behaviour(keypair),
            })
        }
@@ -1,10 +1,11 @@
 { inputs, ... }:
 {
  perSystem =
-    { inputs', pkgs, lib, ... }:
+    { config, self', inputs', pkgs, lib, ... }:
    let
      # Fenix nightly toolchain with all components
-      rustToolchain = inputs'.fenix.packages.stable.withComponents [
+      fenixPkgs = inputs'.fenix.packages;
+      rustToolchain = fenixPkgs.complete.withComponents [
        "cargo"
        "rustc"
        "clippy"
@@ -0,0 +1,2 @@
+[toolchain]
+channel = "nightly"
@@ -1,7 +1,7 @@
 import asyncio
 import socket
 from dataclasses import dataclass, field
-from random import random
+from typing import Iterator

 import anyio
 from anyio import current_time
@@ -12,24 +12,19 @@ from exo.download.download_utils import (
    RepoDownloadProgress,
    delete_model,
    map_repo_download_progress_to_download_progress_data,
-    resolve_model_in_path,
 )
 from exo.download.shard_downloader import ShardDownloader
-from exo.shared.constants import EXO_MODELS_DIR, EXO_MODELS_PATH
-from exo.shared.models.model_cards import ModelId, get_model_cards
+from exo.shared.models.model_cards import ModelId
 from exo.shared.types.commands import (
    CancelDownload,
    DeleteDownload,
    ForwarderDownloadCommand,
    StartDownload,
 )
-from exo.shared.types.common import NodeId, SessionId, SystemId
+from exo.shared.types.common import NodeId, SessionId
 from exo.shared.types.events import (
    Event,
-    EventId,
-    # TODO(evan): just for acks, should delete this ASAP
-    GlobalForwarderEvent,
-    LocalForwarderEvent,
+    ForwarderEvent,
    NodeDownloadProgress,
 )
 from exo.shared.types.worker.downloads import (
@@ -39,7 +34,7 @@ from exo.shared.types.worker.downloads import (
    DownloadPending,
    DownloadProgress,
 )
-from exo.shared.types.worker.shards import PipelineShardMetadata, ShardMetadata
+from exo.shared.types.worker.shards import ShardMetadata
 from exo.utils.channels import Receiver, Sender, channel


@@ -49,15 +44,8 @@ class DownloadCoordinator:
    session_id: SessionId
    shard_downloader: ShardDownloader
    download_command_receiver: Receiver[ForwarderDownloadCommand]
-    local_event_sender: Sender[LocalForwarderEvent]
-
-    # ack stuff
-    _global_event_receiver: Receiver[GlobalForwarderEvent]
-    _out_for_delivery: dict[EventId, LocalForwarderEvent] = field(default_factory=dict)
-
-    offline: bool = False
-
-    _system_id: SystemId = field(default_factory=SystemId)
+    local_event_sender: Sender[ForwarderEvent]
+    event_index_counter: Iterator[int]

    # Local state
    download_status: dict[ModelId, DownloadProgress] = field(default_factory=dict)
@@ -73,13 +61,8 @@ class DownloadCoordinator:

    def __post_init__(self) -> None:
        self.event_sender, self.event_receiver = channel[Event]()
-        if self.offline:
-            self.shard_downloader.set_internet_connection(False)
        self.shard_downloader.on_progress(self._download_progress_callback)

-    def _model_dir(self, model_id: ModelId) -> str:
-        return str(EXO_MODELS_DIR / model_id.normalize())
-
    async def _download_progress_callback(
        self, callback_shard: ShardMetadata, progress: RepoDownloadProgress
    ) -> None:
@@ -90,8 +73,7 @@ class DownloadCoordinator:
            completed = DownloadCompleted(
                shard_metadata=callback_shard,
                node_id=self.node_id,
-                total=progress.total,
-                model_directory=self._model_dir(model_id),
+                total_bytes=progress.total_bytes,
            )
            self.download_status[model_id] = completed
            await self.event_sender.send(
@@ -111,7 +93,6 @@ class DownloadCoordinator:
                download_progress=map_repo_download_progress_to_download_progress_data(
                    progress
                ),
-                model_directory=self._model_dir(model_id),
            )
            self.download_status[model_id] = ongoing
            await self.event_sender.send(
@@ -120,36 +101,23 @@ class DownloadCoordinator:
            self._last_progress_time[model_id] = current_time()

    async def run(self) -> None:
-        logger.info(
-            f"Starting DownloadCoordinator{' (offline mode)' if self.offline else ''}"
-        )
-        if not self.offline:
-            self._test_internet_connection()
-        try:
-            async with self._tg as tg:
-                tg.start_soon(self._command_processor)
-                tg.start_soon(self._forward_events)
-                tg.start_soon(self._emit_existing_download_progress)
-                tg.start_soon(self._resend_out_for_delivery)
-                tg.start_soon(self._clear_ofd)
-                if not self.offline:
-                    tg.start_soon(self._check_internet_connection)
-        finally:
-            for task in self.active_downloads.values():
-                task.cancel()
+        logger.info("Starting DownloadCoordinator")
+        self._test_internet_connection()
+        async with self._tg as tg:
+            tg.start_soon(self._command_processor)
+            tg.start_soon(self._forward_events)
+            tg.start_soon(self._emit_existing_download_progress)
+            tg.start_soon(self._check_internet_connection)

    def _test_internet_connection(self) -> None:
-        # Try multiple endpoints since some ISPs/networks block specific IPs
-        for host in ("1.1.1.1", "8.8.8.8", "1.0.0.1"):
-            try:
-                socket.create_connection((host, 443), timeout=3).close()
-                self.shard_downloader.set_internet_connection(True)
-                logger.debug(f"Internet connectivity: True (via {host})")
-                return
-            except OSError:
-                continue
-        self.shard_downloader.set_internet_connection(False)
-        logger.debug("Internet connectivity: False")
+        try:
+            socket.create_connection(("1.1.1.1", 443), timeout=3).close()
+            self.shard_downloader.set_internet_connection(True)
+        except OSError:
+            self.shard_downloader.set_internet_connection(False)
+        logger.debug(
+            f"Internet connectivity: {self.shard_downloader.internet_connection}"
+        )

    async def _check_internet_connection(self) -> None:
        first_connection = True
@@ -169,20 +137,6 @@ class DownloadCoordinator:
    def shutdown(self) -> None:
        self._tg.cancel_scope.cancel()

-    # directly copied from worker
-    async def _resend_out_for_delivery(self) -> None:
-        # This can also be massively tightened, we should check events are at least a certain age before resending.
-        # Exponential backoff would also certainly help here.
-        while True:
-            await anyio.sleep(1 + random())
-            for event in self._out_for_delivery.copy().values():
-                await self.local_event_sender.send(event)
-
-    async def _clear_ofd(self) -> None:
-        with self._global_event_receiver as events:
-            async for event in events:
-                self._out_for_delivery.pop(event.event.event_id, None)
-
    async def _command_processor(self) -> None:
        with self.download_command_receiver as commands:
            async for cmd in commands:
@@ -215,31 +169,8 @@ class DownloadCoordinator:
                )
                return

-        # Check EXO_MODELS_PATH for pre-downloaded models
-        found_path = resolve_model_in_path(model_id)
-        if found_path is not None:
-            logger.info(
-                f"DownloadCoordinator: Model {model_id} found in EXO_MODELS_PATH at {found_path}"
-            )
-            completed = DownloadCompleted(
-                shard_metadata=shard,
-                node_id=self.node_id,
-                total=shard.model_card.storage_size,
-                model_directory=str(found_path),
-                read_only=True,
-            )
-            self.download_status[model_id] = completed
-            await self.event_sender.send(
-                NodeDownloadProgress(download_progress=completed)
-            )
-            return
-
        # Emit pending status
-        progress = DownloadPending(
-            shard_metadata=shard,
-            node_id=self.node_id,
-            model_directory=self._model_dir(model_id),
-        )
+        progress = DownloadPending(shard_metadata=shard, node_id=self.node_id)
        self.download_status[model_id] = progress
        await self.event_sender.send(NodeDownloadProgress(download_progress=progress))

@@ -252,8 +183,7 @@ class DownloadCoordinator:
            completed = DownloadCompleted(
                shard_metadata=shard,
                node_id=self.node_id,
-                total=initial_progress.total,
-                model_directory=self._model_dir(model_id),
+                total_bytes=initial_progress.total_bytes,
            )
            self.download_status[model_id] = completed
            await self.event_sender.send(
@@ -261,20 +191,6 @@ class DownloadCoordinator:
            )
            return

-        if self.offline:
-            logger.warning(
-                f"Offline mode: model {model_id} is not fully available locally, cannot download"
-            )
-            failed = DownloadFailed(
-                shard_metadata=shard,
-                node_id=self.node_id,
-                error_message=f"Model files not found locally in offline mode: {model_id}",
-                model_directory=self._model_dir(model_id),
-            )
-            self.download_status[model_id] = failed
-            await self.event_sender.send(NodeDownloadProgress(download_progress=failed))
-            return
-
        # Start actual download
        self._start_download_task(shard, initial_progress)

@@ -290,7 +206,6 @@ class DownloadCoordinator:
            download_progress=map_repo_download_progress_to_download_progress_data(
                initial_progress
            ),
-            model_directory=self._model_dir(model_id),
        )
        self.download_status[model_id] = status
        self.event_sender.send_nowait(NodeDownloadProgress(download_progress=status))
@@ -304,7 +219,6 @@ class DownloadCoordinator:
                    shard_metadata=shard,
                    node_id=self.node_id,
                    error_message=str(e),
-                    model_directory=self._model_dir(model_id),
                )
                self.download_status[model_id] = failed
                await self.event_sender.send(
@@ -318,15 +232,6 @@ class DownloadCoordinator:
        self.active_downloads[model_id] = task

    async def _delete_download(self, model_id: ModelId) -> None:
-        # Protect read-only models (from EXO_MODELS_PATH) from deletion
-        if model_id in self.download_status:
-            current = self.download_status[model_id]
-            if isinstance(current, DownloadCompleted) and current.read_only:
-                logger.warning(
-                    f"Refusing to delete read-only model {model_id} (from EXO_MODELS_PATH)"
-                )
-                return
-
        # Cancel if active
        if model_id in self.active_downloads:
            logger.info(f"Cancelling active download for {model_id} before deletion")
@@ -348,7 +253,6 @@ class DownloadCoordinator:
            pending = DownloadPending(
                shard_metadata=current_status.shard_metadata,
                node_id=self.node_id,
-                model_directory=self._model_dir(model_id),
            )
            await self.event_sender.send(
                NodeDownloadProgress(download_progress=pending)
@@ -356,21 +260,19 @@ class DownloadCoordinator:
            del self.download_status[model_id]

    async def _forward_events(self) -> None:
-        idx = 0
        with self.event_receiver as events:
            async for event in events:
-                fe = LocalForwarderEvent(
+                idx = next(self.event_index_counter)
+                fe = ForwarderEvent(
                    origin_idx=idx,
-                    origin=self._system_id,
+                    origin=self.node_id,
                    session=self.session_id,
                    event=event,
                )
-                idx += 1
                logger.debug(
                    f"DownloadCoordinator published event {idx}: {str(event)[:100]}"
                )
                await self.local_event_sender.send(fe)
-                self._out_for_delivery[event.event_id] = fe

    async def _emit_existing_download_progress(self) -> None:
        try:
@@ -392,21 +294,22 @@ class DownloadCoordinator:
                        status: DownloadProgress = DownloadCompleted(
                            node_id=self.node_id,
                            shard_metadata=progress.shard,
-                            total=progress.total,
-                            model_directory=self._model_dir(
-                                progress.shard.model_card.model_id
-                            ),
+                            total_bytes=progress.total_bytes,
                        )
                    elif progress.status in ["in_progress", "not_started"]:
-                        if progress.downloaded_this_session.in_bytes == 0:
-                            status = DownloadPending(
+                        if (
+                            progress.downloaded_bytes.in_bytes
+                            >= progress.total_bytes.in_bytes
+                            > 0
+                        ):
+                            status = DownloadCompleted(
                                node_id=self.node_id,
                                shard_metadata=progress.shard,
-                                model_directory=self._model_dir(
-                                    progress.shard.model_card.model_id
-                                ),
-                                downloaded=progress.downloaded,
-                                total=progress.total,
+                                total_bytes=progress.total_bytes,
+                            )
+                        elif progress.downloaded_bytes_this_session.in_bytes == 0:
+                            status = DownloadPending(
+                                node_id=self.node_id, shard_metadata=progress.shard
                            )
                        else:
                            status = DownloadOngoing(
@@ -415,9 +318,6 @@ class DownloadCoordinator:
                                download_progress=map_repo_download_progress_to_download_progress_data(
                                    progress
                                ),
-                                model_directory=self._model_dir(
-                                    progress.shard.model_card.model_id
-                                ),
                            )
                    else:
                        continue
@@ -426,39 +326,6 @@ class DownloadCoordinator:
                    await self.event_sender.send(
                        NodeDownloadProgress(download_progress=status)
                    )
-                # Scan EXO_MODELS_PATH for pre-downloaded models
-                if EXO_MODELS_PATH is not None:
-                    for card in await get_model_cards():
-                        mid = card.model_id
-                        if mid in self.active_downloads:
-                            continue
-                        if isinstance(
-                            self.download_status.get(mid),
-                            (DownloadCompleted, DownloadOngoing, DownloadFailed),
-                        ):
-                            continue
-                        found = resolve_model_in_path(mid)
-                        if found is not None:
-                            path_shard = PipelineShardMetadata(
-                                model_card=card,
-                                device_rank=0,
-                                world_size=1,
-                                start_layer=0,
-                                end_layer=card.n_layers,
-                                n_layers=card.n_layers,
-                            )
-                            path_completed: DownloadProgress = DownloadCompleted(
-                                node_id=self.node_id,
-                                shard_metadata=path_shard,
-                                total=card.storage_size,
-                                model_directory=str(found),
-                                read_only=True,
-                            )
-                            self.download_status[mid] = path_completed
-                            await self.event_sender.send(
-                                NodeDownloadProgress(download_progress=path_completed)
-                            )
-
                logger.debug(
                    "DownloadCoordinator: Done emitting existing download progress."
                )
@@ -20,6 +20,7 @@ from huggingface_hub import (
 )
 from loguru import logger
 from pydantic import (
+    DirectoryPath,
    TypeAdapter,
 )

@@ -30,7 +31,7 @@ from exo.download.huggingface_utils import (
    get_hf_endpoint,
    get_hf_token,
 )
-from exo.shared.constants import EXO_MODELS_DIR, EXO_MODELS_PATH
+from exo.shared.constants import EXO_MODELS_DIR
 from exo.shared.models.model_cards import ModelTask
 from exo.shared.types.common import ModelId
 from exo.shared.types.memory import Memory
@@ -79,9 +80,9 @@ def map_repo_file_download_progress_to_download_progress_data(
    repo_file_download_progress: RepoFileDownloadProgress,
 ) -> DownloadProgressData:
    return DownloadProgressData(
-        downloaded=repo_file_download_progress.downloaded,
-        downloaded_this_session=repo_file_download_progress.downloaded_this_session,
-        total=repo_file_download_progress.total,
+        downloaded_bytes=repo_file_download_progress.downloaded,
+        downloaded_bytes_this_session=repo_file_download_progress.downloaded_this_session,
+        total_bytes=repo_file_download_progress.total,
        completed_files=1 if repo_file_download_progress.status == "complete" else 0,
        total_files=1,
        speed=repo_file_download_progress.speed,
@@ -94,9 +95,9 @@ def map_repo_download_progress_to_download_progress_data(
    repo_download_progress: RepoDownloadProgress,
 ) -> DownloadProgressData:
    return DownloadProgressData(
-        total=repo_download_progress.total,
-        downloaded=repo_download_progress.downloaded,
-        downloaded_this_session=repo_download_progress.downloaded_this_session,
+        total_bytes=repo_download_progress.total_bytes,
+        downloaded_bytes=repo_download_progress.downloaded_bytes,
+        downloaded_bytes_this_session=repo_download_progress.downloaded_bytes_this_session,
        completed_files=repo_download_progress.completed_files,
        total_files=repo_download_progress.total_files,
        speed=repo_download_progress.overall_speed,
@@ -110,27 +111,7 @@ def map_repo_download_progress_to_download_progress_data(
    )


-def resolve_model_in_path(model_id: ModelId) -> Path | None:
-    """Search EXO_MODELS_PATH directories for a pre-existing model.
-
-    Checks each directory for the normalized name (org--model).  A candidate
-    is only returned if ``is_model_directory_complete`` confirms all weight
-    files are present.
-    """
-    if EXO_MODELS_PATH is None:
-        return None
-    normalized = model_id.normalize()
-    for search_dir in EXO_MODELS_PATH:
-        candidate = search_dir / normalized
-        if candidate.is_dir() and is_model_directory_complete(candidate):
-            return candidate
-    return None
-
-
-def build_model_path(model_id: ModelId) -> Path:
-    found = resolve_model_in_path(model_id)
-    if found is not None:
-        return found
+def build_model_path(model_id: ModelId) -> DirectoryPath:
    return EXO_MODELS_DIR / model_id.normalize()


@@ -161,7 +142,7 @@ async def delete_model(model_id: ModelId) -> bool:


 async def seed_models(seed_dir: str | Path):
-    """Move models from resources folder to EXO_MODELS_DIR."""
+    """Move model in resources folder of app to .cache/huggingface/hub"""
    source_dir = Path(seed_dir)
    dest_dir = await ensure_models_dir()
    for path in source_dir.iterdir():
@@ -177,72 +158,6 @@ async def seed_models(seed_dir: str | Path):
                    logger.error(traceback.format_exc())


-def _scan_model_directory(
-    model_dir: Path, recursive: bool = False
-) -> list[FileListEntry] | None:
-    """Scan a local model directory and build a file list.
-
-    Requires at least one ``*.safetensors.index.json``.  Every weight file
-    referenced by the index that is missing on disk gets ``size=None``.
-    """
-    index_files = list(model_dir.glob("**/*.safetensors.index.json"))
-    if not index_files:
-        return None
-
-    entries_by_path: dict[str, FileListEntry] = {}
-
-    if recursive:
-        for dirpath, _, filenames in os.walk(model_dir):
-            for filename in filenames:
-                if filename.endswith(".partial"):
-                    continue
-                full_path = Path(dirpath) / filename
-                rel_path = str(full_path.relative_to(model_dir))
-                entries_by_path[rel_path] = FileListEntry(
-                    type="file",
-                    path=rel_path,
-                    size=full_path.stat().st_size,
-                )
-    else:
-        for item in model_dir.iterdir():
-            if item.is_file() and not item.name.endswith(".partial"):
-                entries_by_path[item.name] = FileListEntry(
-                    type="file",
-                    path=item.name,
-                    size=item.stat().st_size,
-                )
-
-    # Add expected weight files from index that haven't been downloaded yet
-    for index_file in index_files:
-        try:
-            index_data = ModelSafetensorsIndex.model_validate_json(
-                index_file.read_text()
-            )
-            relative_dir = index_file.parent.relative_to(model_dir)
-            for filename in set(index_data.weight_map.values()):
-                rel_path = (
-                    str(relative_dir / filename)
-                    if relative_dir != Path(".")
-                    else filename
-                )
-                if rel_path not in entries_by_path:
-                    entries_by_path[rel_path] = FileListEntry(
-                        type="file",
-                        path=rel_path,
-                        size=None,
-                    )
-        except Exception:
-            continue
-
-    return list(entries_by_path.values())
-
-
-def is_model_directory_complete(model_dir: Path) -> bool:
-    """Check if a model directory contains all required weight files."""
-    file_list = _scan_model_directory(model_dir, recursive=True)
-    return file_list is not None and all(f.size is not None for f in file_list)
-
-
 async def _build_file_list_from_local_directory(
    model_id: ModelId,
    recursive: bool = False,
@@ -257,7 +172,59 @@ async def _build_file_list_from_local_directory(
    if not await aios.path.exists(model_dir):
        return None

-    file_list = await asyncio.to_thread(_scan_model_directory, model_dir, recursive)
+    def _scan() -> list[FileListEntry] | None:
+        index_files = list(model_dir.glob("**/*.safetensors.index.json"))
+        if not index_files:
+            return None
+
+        entries_by_path: dict[str, FileListEntry] = {}
+
+        if recursive:
+            for dirpath, _, filenames in os.walk(model_dir):
+                for filename in filenames:
+                    if filename.endswith(".partial"):
+                        continue
+                    full_path = Path(dirpath) / filename
+                    rel_path = str(full_path.relative_to(model_dir))
+                    entries_by_path[rel_path] = FileListEntry(
+                        type="file",
+                        path=rel_path,
+                        size=full_path.stat().st_size,
+                    )
+        else:
+            for item in model_dir.iterdir():
+                if item.is_file() and not item.name.endswith(".partial"):
+                    entries_by_path[item.name] = FileListEntry(
+                        type="file",
+                        path=item.name,
+                        size=item.stat().st_size,
+                    )
+
+        # Add expected weight files from index that haven't been downloaded yet
+        for index_file in index_files:
+            try:
+                index_data = ModelSafetensorsIndex.model_validate_json(
+                    index_file.read_text()
+                )
+                relative_dir = index_file.parent.relative_to(model_dir)
+                for filename in set(index_data.weight_map.values()):
+                    rel_path = (
+                        str(relative_dir / filename)
+                        if relative_dir != Path(".")
+                        else filename
+                    )
+                    if rel_path not in entries_by_path:
+                        entries_by_path[rel_path] = FileListEntry(
+                            type="file",
+                            path=rel_path,
+                            size=None,
+                        )
+            except Exception:
+                continue
+
+        return list(entries_by_path.values())
+
+    file_list = await asyncio.to_thread(_scan)
    if not file_list:
        return None
    return file_list
@@ -481,13 +448,12 @@ async def download_file_with_retry(
    target_dir: Path,
    on_progress: Callable[[int, int, bool], None] = lambda _, __, ___: None,
    on_connection_lost: Callable[[], None] = lambda: None,
-    skip_internet: bool = False,
 ) -> Path:
    n_attempts = 3
    for attempt in range(n_attempts):
        try:
            return await _download_file(
-                model_id, revision, path, target_dir, on_progress, skip_internet
+                model_id, revision, path, target_dir, on_progress
            )
        except HuggingFaceAuthenticationError:
            raise
@@ -521,14 +487,10 @@ async def _download_file(
    path: str,
    target_dir: Path,
    on_progress: Callable[[int, int, bool], None] = lambda _, __, ___: None,
-    skip_internet: bool = False,
 ) -> Path:
    target_path = target_dir / path

    if await aios.path.exists(target_path):
-        if skip_internet:
-            return target_path
-
        local_size = (await aios.stat(target_path)).st_size

        # Try to verify against remote, but allow offline operation
@@ -548,11 +510,6 @@ async def _download_file(
            )
            return target_path

-    if skip_internet:
-        raise FileNotFoundError(
-            f"File {path} not found locally and cannot download in offline mode"
-        )
-
    await aios.makedirs((target_dir / path).parent, exist_ok=True)
    length, etag = await file_meta(model_id, revision, path)
    remote_hash = etag[:-5] if etag.endswith("-gzip") else etag
@@ -611,20 +568,19 @@ def calculate_repo_progress(
    file_progress: dict[str, RepoFileDownloadProgress],
    all_start_time: float,
 ) -> RepoDownloadProgress:
-    all_total = sum((p.total for p in file_progress.values()), Memory.from_bytes(0))
-    all_downloaded = sum(
-        (p.downloaded for p in file_progress.values()), Memory.from_bytes(0)
+    all_total_bytes = sum((p.total.in_bytes for p in file_progress.values()), 0)
+    all_downloaded_bytes = sum(
+        (p.downloaded.in_bytes for p in file_progress.values()), 0
    )
-    all_downloaded_this_session = sum(
-        (p.downloaded_this_session for p in file_progress.values()),
-        Memory.from_bytes(0),
+    all_downloaded_bytes_this_session = sum(
+        (p.downloaded_this_session.in_bytes for p in file_progress.values()), 0
    )
    elapsed_time = time.time() - all_start_time
    all_speed = (
-        all_downloaded_this_session.in_bytes / elapsed_time if elapsed_time > 0 else 0
+        all_downloaded_bytes_this_session / elapsed_time if elapsed_time > 0 else 0
    )
    all_eta = (
-        timedelta(seconds=(all_total - all_downloaded).in_bytes / all_speed)
+        timedelta(seconds=(all_total_bytes - all_downloaded_bytes) / all_speed)
        if all_speed > 0
        else timedelta(seconds=0)
    )
@@ -643,9 +599,11 @@ def calculate_repo_progress(
            [p for p in file_progress.values() if p.downloaded == p.total]
        ),
        total_files=len(file_progress),
-        downloaded=all_downloaded,
-        downloaded_this_session=all_downloaded_this_session,
-        total=all_total,
+        downloaded_bytes=Memory.from_bytes(all_downloaded_bytes),
+        downloaded_bytes_this_session=Memory.from_bytes(
+            all_downloaded_bytes_this_session
+        ),
+        total_bytes=Memory.from_bytes(all_total_bytes),
        overall_speed=all_speed,
        overall_eta=all_eta,
        status=status,
@@ -856,7 +814,6 @@ async def download_shard(
                    file, curr_bytes, total_bytes, is_renamed
                ),
                on_connection_lost=on_connection_lost,
-                skip_internet=skip_internet,
            )

    if not skip_download:
@@ -107,9 +107,9 @@ NOOP_DOWNLOAD_PROGRESS = RepoDownloadProgress(
    ),
    completed_files=0,
    total_files=0,
-    downloaded=Memory.from_bytes(0),
-    downloaded_this_session=Memory.from_bytes(0),
-    total=Memory.from_bytes(0),
+    downloaded_bytes=Memory.from_bytes(0),
+    downloaded_bytes_this_session=Memory.from_bytes(0),
+    total_bytes=Memory.from_bytes(0),
    overall_speed=0,
    overall_eta=timedelta(seconds=0),
    status="complete",
@@ -1,98 +0,0 @@
-from typing import Any
-
-import anyio
-import pytest
-
-from exo.download.coordinator import DownloadCoordinator
-from exo.download.shard_downloader import NoopShardDownloader
-from exo.shared.models.model_cards import ModelCard, ModelTask
-from exo.shared.types.common import ModelId, NodeId, SessionId
-from exo.shared.types.events import (
-    GlobalForwarderEvent,
-    LocalForwarderEvent,
-    NodeDownloadProgress,
-)
-from exo.shared.types.memory import Memory
-from exo.shared.types.worker.downloads import (
-    DownloadPending,
-)
-from exo.shared.types.worker.shards import PipelineShardMetadata
-from exo.utils.channels import channel
-
-# Use the built‑in NoopShardDownloader directly – it already implements the required abstract interface.
-# No additional subclass is needed for this test.
-
-
-@pytest.mark.anyio
-async def test_ack_behaviour():
-    # Create channels (type Any for simplicity)
-    _, command_receiver = channel[Any]()
-    local_sender, _ = channel[Any]()
-    global_sender, global_receiver = channel[Any]()
-
-    # Minimal identifiers
-    node_id = NodeId()
-    session_id = SessionId(master_node_id=node_id, election_clock=0)
-
-    # Create a dummy model card and shard metadata
-    model_id = ModelId("test/model")
-    model_card = ModelCard(
-        model_id=model_id,
-        storage_size=Memory.from_bytes(0),
-        n_layers=1,
-        hidden_size=1,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    )
-    shard = PipelineShardMetadata(
-        model_card=model_card,
-        device_rank=0,
-        world_size=1,
-        start_layer=0,
-        end_layer=1,
-        n_layers=1,
-    )
-
-    # Instantiate the coordinator with the dummy downloader
-    coord = DownloadCoordinator(
-        node_id=node_id,
-        session_id=session_id,
-        shard_downloader=NoopShardDownloader(),
-        download_command_receiver=command_receiver,
-        local_event_sender=local_sender,
-        _global_event_receiver=global_receiver,
-    )
-
-    async with anyio.create_task_group() as tg:
-        # Start the forwarding and ack‑clearing loops
-        tg.start_soon(coord._forward_events)  # pyright: ignore[reportPrivateUsage]
-        tg.start_soon(coord._clear_ofd)  # pyright: ignore[reportPrivateUsage]
-
-        # Send a pending download progress event via the internal event sender
-        pending = DownloadPending(
-            node_id=node_id,
-            shard_metadata=shard,
-            model_directory="/tmp/model",
-        )
-        await coord.event_sender.send(NodeDownloadProgress(download_progress=pending))
-        # Allow the forwarder to process the event
-        await anyio.sleep(0.1)
-
-        # There should be exactly one entry awaiting ACK
-        assert len(coord._out_for_delivery) == 1  # pyright: ignore[reportPrivateUsage]
-        # Retrieve the stored LocalForwarderEvent
-        stored_fe: LocalForwarderEvent = next(iter(coord._out_for_delivery.values()))  # pyright: ignore[reportPrivateUsage]
-        # Simulate receiving a global ack for this event
-        ack = GlobalForwarderEvent(
-            origin_idx=0,
-            origin=node_id,
-            session=session_id,
-            event=stored_fe.event,
-        )
-        await global_sender.send(ack)
-        # Give the clear‑ofd task a moment to process the ack
-        await anyio.sleep(0.1)
-        # The out‑for‑delivery map should now be empty
-        assert len(coord._out_for_delivery) == 0  # pyright: ignore[reportPrivateUsage]
-        # Cancel background tasks
-        tg.cancel_scope.cancel()
@@ -1,230 +0,0 @@
-"""Tests for offline/air-gapped mode."""
-
-from collections.abc import AsyncIterator
-from pathlib import Path
-from unittest.mock import AsyncMock, patch
-
-import aiofiles
-import aiofiles.os as aios
-import pytest
-
-from exo.download.download_utils import (
-    _download_file,  # pyright: ignore[reportPrivateUsage]
-    download_file_with_retry,
-    fetch_file_list_with_cache,
-)
-from exo.shared.types.common import ModelId
-from exo.shared.types.worker.downloads import FileListEntry
-
-
-@pytest.fixture
-def model_id() -> ModelId:
-    return ModelId("test-org/test-model")
-
-
-@pytest.fixture
-async def temp_models_dir(tmp_path: Path) -> AsyncIterator[Path]:
-    models_dir = tmp_path / "models"
-    await aios.makedirs(models_dir, exist_ok=True)
-    with patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir):
-        yield models_dir
-
-
-class TestDownloadFileOffline:
-    """Tests for _download_file with skip_internet=True."""
-
-    async def test_returns_local_file_without_http_verification(
-        self, model_id: ModelId, tmp_path: Path
-    ) -> None:
-        """When skip_internet=True and file exists locally, return it immediately
-        without making any HTTP calls (no file_meta verification)."""
-        target_dir = tmp_path / "downloads"
-        await aios.makedirs(target_dir, exist_ok=True)
-
-        local_file = target_dir / "model.safetensors"
-        async with aiofiles.open(local_file, "wb") as f:
-            await f.write(b"model weights data")
-
-        with patch(
-            "exo.download.download_utils.file_meta",
-            new_callable=AsyncMock,
-        ) as mock_file_meta:
-            result = await _download_file(
-                model_id,
-                "main",
-                "model.safetensors",
-                target_dir,
-                skip_internet=True,
-            )
-
-            assert result == local_file
-            mock_file_meta.assert_not_called()
-
-    async def test_raises_file_not_found_for_missing_file(
-        self, model_id: ModelId, tmp_path: Path
-    ) -> None:
-        """When skip_internet=True and file does NOT exist locally,
-        raise FileNotFoundError instead of attempting download."""
-        target_dir = tmp_path / "downloads"
-        await aios.makedirs(target_dir, exist_ok=True)
-
-        with pytest.raises(FileNotFoundError, match="offline mode"):
-            await _download_file(
-                model_id,
-                "main",
-                "missing_model.safetensors",
-                target_dir,
-                skip_internet=True,
-            )
-
-    async def test_returns_local_file_in_subdirectory(
-        self, model_id: ModelId, tmp_path: Path
-    ) -> None:
-        """When skip_internet=True and file exists in a subdirectory,
-        return it without HTTP calls."""
-        target_dir = tmp_path / "downloads"
-        subdir = target_dir / "transformer"
-        await aios.makedirs(subdir, exist_ok=True)
-
-        local_file = subdir / "diffusion_pytorch_model.safetensors"
-        async with aiofiles.open(local_file, "wb") as f:
-            await f.write(b"weights")
-
-        with patch(
-            "exo.download.download_utils.file_meta",
-            new_callable=AsyncMock,
-        ) as mock_file_meta:
-            result = await _download_file(
-                model_id,
-                "main",
-                "transformer/diffusion_pytorch_model.safetensors",
-                target_dir,
-                skip_internet=True,
-            )
-
-            assert result == local_file
-            mock_file_meta.assert_not_called()
-
-
-class TestDownloadFileWithRetryOffline:
-    """Tests for download_file_with_retry with skip_internet=True."""
-
-    async def test_propagates_skip_internet_to_download_file(
-        self, model_id: ModelId, tmp_path: Path
-    ) -> None:
-        """Verify skip_internet is passed through to _download_file."""
-        target_dir = tmp_path / "downloads"
-        await aios.makedirs(target_dir, exist_ok=True)
-
-        local_file = target_dir / "config.json"
-        async with aiofiles.open(local_file, "wb") as f:
-            await f.write(b'{"model_type": "qwen2"}')
-
-        with patch(
-            "exo.download.download_utils.file_meta",
-            new_callable=AsyncMock,
-        ) as mock_file_meta:
-            result = await download_file_with_retry(
-                model_id,
-                "main",
-                "config.json",
-                target_dir,
-                skip_internet=True,
-            )
-
-            assert result == local_file
-            mock_file_meta.assert_not_called()
-
-    async def test_file_not_found_does_not_retry(
-        self, model_id: ModelId, tmp_path: Path
-    ) -> None:
-        """FileNotFoundError from offline mode should not trigger retries."""
-        target_dir = tmp_path / "downloads"
-        await aios.makedirs(target_dir, exist_ok=True)
-
-        with pytest.raises(FileNotFoundError):
-            await download_file_with_retry(
-                model_id,
-                "main",
-                "nonexistent.safetensors",
-                target_dir,
-                skip_internet=True,
-            )
-
-
-class TestFetchFileListOffline:
-    """Tests for fetch_file_list_with_cache with skip_internet=True."""
-
-    async def test_uses_cached_file_list(
-        self, model_id: ModelId, temp_models_dir: Path
-    ) -> None:
-        """When skip_internet=True and cache file exists, use it without network."""
-        from pydantic import TypeAdapter
-
-        cache_dir = temp_models_dir / "caches" / model_id.normalize()
-        await aios.makedirs(cache_dir, exist_ok=True)
-
-        cached_list = [
-            FileListEntry(type="file", path="model.safetensors", size=1000),
-            FileListEntry(type="file", path="config.json", size=200),
-        ]
-        cache_file = cache_dir / f"{model_id.normalize()}--main--file_list.json"
-        async with aiofiles.open(cache_file, "w") as f:
-            await f.write(
-                TypeAdapter(list[FileListEntry]).dump_json(cached_list).decode()
-            )
-
-        with patch(
-            "exo.download.download_utils.fetch_file_list_with_retry",
-            new_callable=AsyncMock,
-        ) as mock_fetch:
-            result = await fetch_file_list_with_cache(
-                model_id, "main", skip_internet=True
-            )
-
-            assert result == cached_list
-            mock_fetch.assert_not_called()
-
-    async def test_falls_back_to_local_directory_scan(
-        self, model_id: ModelId, temp_models_dir: Path
-    ) -> None:
-        """When skip_internet=True and no cache but local files exist,
-        build file list from local directory."""
-        import json
-
-        model_dir = temp_models_dir / model_id.normalize()
-        await aios.makedirs(model_dir, exist_ok=True)
-
-        async with aiofiles.open(model_dir / "config.json", "w") as f:
-            await f.write('{"model_type": "qwen2"}')
-
-        index_data = {
-            "metadata": {},
-            "weight_map": {"model.layers.0.weight": "model.safetensors"},
-        }
-        async with aiofiles.open(model_dir / "model.safetensors.index.json", "w") as f:
-            await f.write(json.dumps(index_data))
-
-        async with aiofiles.open(model_dir / "model.safetensors", "wb") as f:
-            await f.write(b"x" * 500)
-
-        with patch(
-            "exo.download.download_utils.fetch_file_list_with_retry",
-            new_callable=AsyncMock,
-        ) as mock_fetch:
-            result = await fetch_file_list_with_cache(
-                model_id, "main", skip_internet=True
-            )
-
-            mock_fetch.assert_not_called()
-            paths = {entry.path for entry in result}
-            assert "config.json" in paths
-            assert "model.safetensors" in paths
-
-    async def test_raises_when_no_cache_and_no_local_files(
-        self, model_id: ModelId, temp_models_dir: Path
-    ) -> None:
-        """When skip_internet=True and neither cache nor local files exist,
-        raise FileNotFoundError."""
-        with pytest.raises(FileNotFoundError, match="No internet"):
-            await fetch_file_list_with_cache(model_id, "main", skip_internet=True)
@@ -1,10 +1,11 @@
 import argparse
+import itertools
 import multiprocessing as mp
 import os
 import resource
 import signal
 from dataclasses import dataclass, field
-from typing import Self
+from typing import Iterator, Self

 import anyio
 from anyio.abc import TaskGroup
@@ -37,15 +38,20 @@ class Node:
    api: API | None

    node_id: NodeId
-    offline: bool
+    event_index_counter: Iterator[int]
    _tg: TaskGroup = field(init=False, default_factory=anyio.create_task_group)

    @classmethod
-    async def create(cls, args: "Args") -> Self:
+    async def create(cls, args: "Args") -> "Self":
        keypair = get_node_id_keypair()
-        node_id = NodeId(keypair.to_node_id())
+        node_id = NodeId(keypair.to_peer_id().to_base58())
        session_id = SessionId(master_node_id=node_id, election_clock=0)
-        router = Router.create(keypair)
+        bootstrap_peers: list[str] = []
+        if args.bootstrap_peer:
+            ip, port = args.bootstrap_peer.rsplit(":", 1)
+            bootstrap_peers.append(f"/ip4/{ip}/tcp/{port}")
+            logger.info(f"Bootstrap peer: {args.bootstrap_peer}")
+        router = Router.create(keypair, bootstrap_peers=bootstrap_peers)
        await router.register_topic(topics.GLOBAL_EVENTS)
        await router.register_topic(topics.LOCAL_EVENTS)
        await router.register_topic(topics.COMMANDS)
@@ -55,6 +61,9 @@ class Node:

        logger.info(f"Starting node {node_id}")

+        # Create shared event index counter for Worker and DownloadCoordinator
+        event_index_counter = itertools.count()
+
        # Create DownloadCoordinator (unless --no-downloads)
        if not args.no_downloads:
            download_coordinator = DownloadCoordinator(
@@ -63,9 +72,7 @@ class Node:
                exo_shard_downloader(),
                download_command_receiver=router.receiver(topics.DOWNLOAD_COMMANDS),
                local_event_sender=router.sender(topics.LOCAL_EVENTS),
-                offline=args.offline,
-                # TODO(evan): remove
-                _global_event_receiver=router.receiver(topics.GLOBAL_EVENTS),
+                event_index_counter=event_index_counter,
            )
        else:
            download_coordinator = None
@@ -91,6 +98,7 @@ class Node:
                local_event_sender=router.sender(topics.LOCAL_EVENTS),
                command_sender=router.sender(topics.COMMANDS),
                download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
+                event_index_counter=event_index_counter,
            )
        else:
            worker = None
@@ -128,13 +136,11 @@ class Node:
            master,
            api,
            node_id,
-            args.offline,
+            event_index_counter,
        )

    async def run(self):
        async with self._tg as tg:
-            signal.signal(signal.SIGINT, lambda _, __: self.shutdown())
-            signal.signal(signal.SIGTERM, lambda _, __: self.shutdown())
            tg.start_soon(self.router.run)
            tg.start_soon(self.election.run)
            if self.download_coordinator:
@@ -146,6 +152,8 @@ class Node:
            if self.api:
                tg.start_soon(self.api.run)
            tg.start_soon(self._elect_loop)
+            signal.signal(signal.SIGINT, lambda _, __: self.shutdown())
+            signal.signal(signal.SIGTERM, lambda _, __: self.shutdown())

    def shutdown(self):
        # if this is our second call to shutdown, just sys.exit
@@ -206,6 +214,8 @@ class Node:
                    )
                if result.is_new_master:
                    await anyio.sleep(0)
+                    # Fresh counter for new session (buffer expects indices from 0)
+                    self.event_index_counter = itertools.count()
                    if self.download_coordinator:
                        self.download_coordinator.shutdown()
                        self.download_coordinator = DownloadCoordinator(
@@ -216,11 +226,7 @@ class Node:
                                topics.DOWNLOAD_COMMANDS
                            ),
                            local_event_sender=self.router.sender(topics.LOCAL_EVENTS),
-                            offline=self.offline,
-                            # TODO(evan): remove
-                            _global_event_receiver=self.router.receiver(
-                                topics.GLOBAL_EVENTS
-                            ),
+                            event_index_counter=self.event_index_counter,
                        )
                        self._tg.start_soon(self.download_coordinator.run)
                    if self.worker:
@@ -237,6 +243,7 @@ class Node:
                            download_command_sender=self.router.sender(
                                topics.DOWNLOAD_COMMANDS
                            ),
+                            event_index_counter=self.event_index_counter,
                        )
                        self._tg.start_soon(self.worker.run)
                    if self.api:
@@ -252,15 +259,12 @@ def main():
    target = min(max(soft, 65535), hard)
    resource.setrlimit(resource.RLIMIT_NOFILE, (target, hard))

-    mp.set_start_method("spawn")
+    mp.set_start_method("spawn", force=True)
    # TODO: Refactor the current verbosity system
    logger_setup(EXO_LOG, args.verbosity)
    logger.info("Starting EXO")
    logger.info(f"EXO_LIBP2P_NAMESPACE: {os.getenv('EXO_LIBP2P_NAMESPACE')}")

-    if args.offline:
-        logger.info("Running in OFFLINE mode — no internet checks, local models only")
-
    # Set FAST_SYNCH override env var for runner subprocesses
    if args.fast_synch is True:
        os.environ["EXO_FAST_SYNCH"] = "on"
@@ -283,8 +287,8 @@ class Args(CamelCaseModel):
    tb_only: bool = False
    no_worker: bool = False
    no_downloads: bool = False
-    offline: bool = False
    fast_synch: bool | None = None  # None = auto, True = force on, False = force off
+    bootstrap_peer: str | None = None

    @classmethod
    def parse(cls) -> Self:
@@ -331,11 +335,6 @@ class Args(CamelCaseModel):
            action="store_true",
            help="Disable the download coordinator (node won't download models)",
        )
-        parser.add_argument(
-            "--offline",
-            action="store_true",
-            help="Run in offline/air-gapped mode: skip internet checks, use only pre-staged local models",
-        )
        fast_synch_group = parser.add_mutually_exclusive_group()
        fast_synch_group.add_argument(
            "--fast-synch",
@@ -350,6 +349,13 @@ class Args(CamelCaseModel):
            dest="fast_synch",
            help="Force MLX FAST_SYNCH off",
        )
+        parser.add_argument(
+            "--bootstrap-peer",
+            type=str,
+            dest="bootstrap_peer",
+            default=None,
+            help="IP:PORT of an existing node to connect to (bypasses mDNS)",
+        )

        args = parser.parse_args()
        return cls(**vars(args))  # pyright: ignore[reportAny] - We are intentionally validating here, we can't do it statically
@@ -19,12 +19,7 @@ from exo.shared.types.api import (
    ToolCall,
    Usage,
 )
-from exo.shared.types.chunks import (
-    ErrorChunk,
-    PrefillProgressChunk,
-    TokenChunk,
-    ToolCallChunk,
-)
+from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
 from exo.shared.types.common import CommandId
 from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams

@@ -59,11 +54,7 @@ def chat_request_to_text_generation(
            chat_template_messages.append({"role": "system", "content": content})
        else:
            # Skip messages with no meaningful content
-            if (
-                msg.content is None
-                and msg.reasoning_content is None
-                and msg.tool_calls is None
-            ):
+            if msg.content is None and msg.thinking is None and msg.tool_calls is None:
                continue

            if msg.role in ("user", "assistant", "developer"):
@@ -115,11 +106,6 @@ def chunk_to_response(
            ]
        )

-    if chunk.is_thinking:
-        delta = ChatCompletionMessage(role="assistant", reasoning_content=chunk.text)
-    else:
-        delta = ChatCompletionMessage(role="assistant", content=chunk.text)
-
    return ChatCompletionResponse(
        id=command_id,
        created=int(time.time()),
@@ -127,7 +113,7 @@ def chunk_to_response(
        choices=[
            StreamingChoiceResponse(
                index=0,
-                delta=delta,
+                delta=ChatCompletionMessage(role="assistant", content=chunk.text),
                logprobs=logprobs,
                finish_reason=chunk.finish_reason,
            )
@@ -137,87 +123,72 @@ def chunk_to_response(

 async def generate_chat_stream(
    command_id: CommandId,
-    chunk_stream: AsyncGenerator[
-        PrefillProgressChunk | ErrorChunk | ToolCallChunk | TokenChunk, None
-    ],
+    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:
-        match chunk:
-            case PrefillProgressChunk():
-                # Use SSE comment so third-party clients ignore it
-                yield f": prefill_progress {chunk.model_dump_json()}\n\n"
-
-            case ErrorChunk():
-                error_response = ErrorResponse(
-                    error=ErrorInfo(
-                        message=chunk.error_message or "Internal server error",
-                        type="InternalServerError",
-                        code=500,
-                    )
+        if isinstance(chunk, ErrorChunk):
+            error_response = ErrorResponse(
+                error=ErrorInfo(
+                    message=chunk.error_message or "Internal server error",
+                    type="InternalServerError",
+                    code=500,
                )
-                yield f"data: {error_response.model_dump_json()}\n\n"
-                yield "data: [DONE]\n\n"
-                return
+            )
+            yield f"data: {error_response.model_dump_json()}\n\n"
+            yield "data: [DONE]\n\n"
+            return

-            case ToolCallChunk():
-                last_usage = chunk.usage or last_usage
+        last_usage = chunk.usage or last_usage

-                tool_call_deltas = [
-                    ToolCall(
-                        id=tool.id,
-                        index=i,
-                        function=tool,
-                    )
-                    for i, tool in enumerate(chunk.tool_calls)
-                ]
-                tool_response = ChatCompletionResponse(
-                    id=command_id,
-                    created=int(time.time()),
-                    model=chunk.model,
-                    choices=[
-                        StreamingChoiceResponse(
-                            index=0,
-                            delta=ChatCompletionMessage(
-                                role="assistant",
-                                tool_calls=tool_call_deltas,
-                            ),
-                            finish_reason="tool_calls",
-                        )
-                    ],
-                    usage=last_usage,
+        if isinstance(chunk, ToolCallChunk):
+            tool_call_deltas = [
+                ToolCall(
+                    id=tool.id,
+                    index=i,
+                    function=tool,
                )
-                yield f"data: {tool_response.model_dump_json()}\n\n"
-                yield "data: [DONE]\n\n"
-                return
-
-            case TokenChunk():
-                last_usage = chunk.usage or last_usage
-
-                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}
+                for i, tool in enumerate(chunk.tool_calls)
+            ]
+            tool_response = ChatCompletionResponse(
+                id=command_id,
+                created=int(time.time()),
+                model=chunk.model,
+                choices=[
+                    StreamingChoiceResponse(
+                        index=0,
+                        delta=ChatCompletionMessage(
+                            role="assistant",
+                            tool_calls=tool_call_deltas,
+                        ),
+                        finish_reason="tool_calls",
                    )
-                yield f"data: {chunk_response.model_dump_json()}\n\n"
+                ],
+                usage=last_usage,
+            )
+            yield f"data: {tool_response.model_dump_json()}\n\n"
+            yield "data: [DONE]\n\n"
+            return

-                if chunk.finish_reason is not None:
-                    yield "data: [DONE]\n\n"
+        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:
+            yield "data: [DONE]\n\n"


 async def collect_chat_response(
    command_id: CommandId,
-    chunk_stream: AsyncGenerator[
-        ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
-    ],
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
 ) -> 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] = []
-    thinking_parts: list[str] = []
    tool_calls: list[ToolCall] = []
    logprobs_content: list[LogprobsContentItem] = []
    model: str | None = None
@@ -226,52 +197,43 @@ async def collect_chat_response(
    last_usage: Usage | None = None

    async for chunk in chunk_stream:
-        match chunk:
-            case PrefillProgressChunk():
-                continue
+        if isinstance(chunk, ErrorChunk):
+            error_message = chunk.error_message or "Internal server error"
+            break

-            case ErrorChunk():
-                error_message = chunk.error_message or "Internal server error"
-                break
+        if model is None:
+            model = chunk.model

-            case TokenChunk():
-                if model is None:
-                    model = chunk.model
-                last_usage = chunk.usage or last_usage
-                if chunk.is_thinking:
-                    thinking_parts.append(chunk.text)
-                else:
-                    text_parts.append(chunk.text)
-                if chunk.logprob is not None:
-                    logprobs_content.append(
-                        LogprobsContentItem(
-                            token=chunk.text,
-                            logprob=chunk.logprob,
-                            top_logprobs=chunk.top_logprobs or [],
-                        )
+        last_usage = chunk.usage or last_usage
+
+        if isinstance(chunk, TokenChunk):
+            text_parts.append(chunk.text)
+            if chunk.logprob is not None:
+                logprobs_content.append(
+                    LogprobsContentItem(
+                        token=chunk.text,
+                        logprob=chunk.logprob,
+                        top_logprobs=chunk.top_logprobs or [],
                    )
-                if chunk.finish_reason is not None:
-                    finish_reason = chunk.finish_reason
-
-            case ToolCallChunk():
-                if model is None:
-                    model = chunk.model
-                last_usage = chunk.usage or last_usage
-                tool_calls.extend(
-                    ToolCall(
-                        id=tool.id,
-                        index=i,
-                        function=tool,
-                    )
-                    for i, tool in enumerate(chunk.tool_calls)
                )
-                finish_reason = chunk.finish_reason
+
+        if isinstance(chunk, ToolCallChunk):
+            tool_calls.extend(
+                ToolCall(
+                    id=tool.id,
+                    index=i,
+                    function=tool,
+                )
+                for i, tool in enumerate(chunk.tool_calls)
+            )
+
+        if chunk.finish_reason is not None:
+            finish_reason = chunk.finish_reason

    if error_message is not None:
        raise ValueError(error_message)

    combined_text = "".join(text_parts)
-    combined_thinking = "".join(thinking_parts) if thinking_parts else None
    assert model is not None

    yield ChatCompletionResponse(
@@ -284,7 +246,6 @@ async def collect_chat_response(
                message=ChatCompletionMessage(
                    role="assistant",
                    content=combined_text,
-                    reasoning_content=combined_thinking,
                    tool_calls=tool_calls if tool_calls else None,
                ),
                logprobs=Logprobs(content=logprobs_content)
@@ -1,17 +1,11 @@
 """Claude Messages API adapter for converting requests/responses."""

 import json
-import re
 from collections.abc import AsyncGenerator
 from typing import Any

 from exo.shared.types.api import FinishReason, Usage
-from exo.shared.types.chunks import (
-    ErrorChunk,
-    PrefillProgressChunk,
-    TokenChunk,
-    ToolCallChunk,
-)
+from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
 from exo.shared.types.claude_api import (
    ClaudeContentBlock,
    ClaudeContentBlockDeltaEvent,
@@ -29,8 +23,6 @@ from exo.shared.types.claude_api import (
    ClaudeStopReason,
    ClaudeTextBlock,
    ClaudeTextDelta,
-    ClaudeThinkingBlock,
-    ClaudeThinkingDelta,
    ClaudeToolResultBlock,
    ClaudeToolUseBlock,
    ClaudeUsage,
@@ -64,22 +56,6 @@ def _extract_tool_result_text(block: ClaudeToolResultBlock) -> str:
    return "".join(sub_block.text for sub_block in block.content)


-# Matches "x-anthropic-billing-header: ...;" (with optional trailing newline)
-# or similar telemetry headers that change every request and break KV prefix caching.
-_VOLATILE_HEADER_RE = re.compile(r"^x-anthropic-[^\n]*;\n?", re.MULTILINE)
-
-
-def _strip_volatile_headers(text: str) -> str:
-    """Remove Anthropic billing/telemetry headers from system prompt text.
-
-    Claude Code prepends headers like 'x-anthropic-billing-header: cc_version=...;
-    cc_entrypoint=...; cch=...;' that contain per-request content hashes. These
-    change every request and break KV prefix caching (the prefix diverges at ~20
-    tokens instead of matching thousands of conversation tokens).
-    """
-    return _VOLATILE_HEADER_RE.sub("", text)
-
-
 def claude_request_to_text_generation(
    request: ClaudeMessagesRequest,
 ) -> TextGenerationTaskParams:
@@ -92,8 +68,6 @@ def claude_request_to_text_generation(
            instructions = request.system
        else:
            instructions = "".join(block.text for block in request.system)
-
-        instructions = _strip_volatile_headers(instructions)
        chat_template_messages.append({"role": "system", "content": instructions})

    # Convert messages to input
@@ -106,15 +80,12 @@ def claude_request_to_text_generation(

        # Process structured content blocks
        text_parts: list[str] = []
-        thinking_parts: list[str] = []
        tool_calls: list[dict[str, Any]] = []
        tool_results: list[ClaudeToolResultBlock] = []

        for block in msg.content:
            if isinstance(block, ClaudeTextBlock):
                text_parts.append(block.text)
-            elif isinstance(block, ClaudeThinkingBlock):
-                thinking_parts.append(block.thinking)
            elif isinstance(block, ClaudeToolUseBlock):
                tool_calls.append(
                    {
@@ -130,7 +101,6 @@ def claude_request_to_text_generation(
                tool_results.append(block)

        content = "".join(text_parts)
-        reasoning_content = "".join(thinking_parts) if thinking_parts else None

        # Build InputMessage from text content
        if msg.role in ("user", "assistant"):
@@ -138,14 +108,9 @@ def claude_request_to_text_generation(

        # Build chat_template_messages preserving tool structure
        if tool_calls:
-            chat_msg: dict[str, Any] = {
-                "role": "assistant",
-                "content": content,
-                "tool_calls": tool_calls,
-            }
-            if reasoning_content:
-                chat_msg["reasoning_content"] = reasoning_content
-            chat_template_messages.append(chat_msg)
+            chat_template_messages.append(
+                {"role": "assistant", "content": content, "tool_calls": tool_calls}
+            )
        elif tool_results:
            for tr in tool_results:
                chat_template_messages.append(
@@ -156,10 +121,7 @@ def claude_request_to_text_generation(
                    }
                )
        else:
-            chat_msg = {"role": msg.role, "content": content}
-            if reasoning_content:
-                chat_msg["reasoning_content"] = reasoning_content
-            chat_template_messages.append(chat_msg)
+            chat_template_messages.append({"role": msg.role, "content": content})

    # Convert Claude tool definitions to OpenAI-style function tools
    tools: list[dict[str, Any]] | None = None
@@ -176,10 +138,6 @@ def claude_request_to_text_generation(
            for tool in request.tools
        ]

-    enable_thinking: bool | None = None
-    if request.thinking is not None:
-        enable_thinking = request.thinking.type in ("enabled", "adaptive")
-
    return TextGenerationTaskParams(
        model=request.model,
        input=input_messages
@@ -193,7 +151,6 @@ def claude_request_to_text_generation(
        stop=request.stop_sequences,
        stream=request.stream,
        tools=tools,
-        enable_thinking=enable_thinking,
        chat_template_messages=chat_template_messages
        if chat_template_messages
        else None,
@@ -203,24 +160,18 @@ def claude_request_to_text_generation(
 async def collect_claude_response(
    command_id: CommandId,
    model: str,
-    chunk_stream: AsyncGenerator[
-        ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
-    ],
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
 ) -> 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] = []
-    thinking_parts: list[str] = []
    tool_use_blocks: list[ClaudeToolUseBlock] = []
    stop_reason: ClaudeStopReason | None = None
    last_usage: Usage | None = None
    error_message: str | None = None

    async for chunk in chunk_stream:
-        if isinstance(chunk, PrefillProgressChunk):
-            continue
-
        if isinstance(chunk, ErrorChunk):
            error_message = chunk.error_message or "Internal server error"
            break
@@ -239,10 +190,7 @@ async def collect_claude_response(
            stop_reason = "tool_use"
            continue

-        if chunk.is_thinking:
-            thinking_parts.append(chunk.text)
-        else:
-            text_parts.append(chunk.text)
+        text_parts.append(chunk.text)

        if chunk.finish_reason is not None:
            stop_reason = finish_reason_to_claude_stop_reason(chunk.finish_reason)
@@ -251,12 +199,9 @@ async def collect_claude_response(
        raise ValueError(error_message)

    combined_text = "".join(text_parts)
-    combined_thinking = "".join(thinking_parts)

    # Build content blocks
    content: list[ClaudeContentBlock] = []
-    if combined_thinking:
-        content.append(ClaudeThinkingBlock(thinking=combined_thinking))
    if combined_text:
        content.append(ClaudeTextBlock(text=combined_text))
    content.extend(tool_use_blocks)
@@ -285,9 +230,7 @@ async def collect_claude_response(
 async def generate_claude_stream(
    command_id: CommandId,
    model: str,
-    chunk_stream: AsyncGenerator[
-        ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
-    ],
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
 ) -> AsyncGenerator[str, None]:
    """Generate Claude Messages API streaming events from TokenChunks."""
    # Initial message_start event
@@ -301,21 +244,18 @@ async def generate_claude_stream(
    start_event = ClaudeMessageStartEvent(message=initial_message)
    yield f"event: message_start\ndata: {start_event.model_dump_json()}\n\n"

+    # content_block_start for text block at index 0
+    block_start = ClaudeContentBlockStartEvent(
+        index=0, content_block=ClaudeTextBlock(text="")
+    )
+    yield f"event: content_block_start\ndata: {block_start.model_dump_json()}\n\n"
+
    output_tokens = 0
    stop_reason: ClaudeStopReason | None = None
    last_usage: Usage | None = None
-    next_block_index = 0
-
-    # Track whether we've started thinking/text blocks
-    thinking_block_started = False
-    thinking_block_index = -1
-    text_block_started = False
-    text_block_index = -1
+    next_block_index = 1  # text block is 0, tool blocks start at 1

    async for chunk in chunk_stream:
-        if isinstance(chunk, PrefillProgressChunk):
-            continue
-
        if isinstance(chunk, ErrorChunk):
            # Close text block and bail
            break
@@ -355,45 +295,12 @@ async def generate_claude_stream(

        output_tokens += 1  # Count each chunk as one token

-        if chunk.is_thinking:
-            # Start thinking block on first thinking token
-            if not thinking_block_started:
-                thinking_block_started = True
-                thinking_block_index = next_block_index
-                next_block_index += 1
-                block_start = ClaudeContentBlockStartEvent(
-                    index=thinking_block_index,
-                    content_block=ClaudeThinkingBlock(thinking=""),
-                )
-                yield f"event: content_block_start\ndata: {block_start.model_dump_json()}\n\n"
-
-            delta_event = ClaudeContentBlockDeltaEvent(
-                index=thinking_block_index,
-                delta=ClaudeThinkingDelta(thinking=chunk.text),
-            )
-            yield f"event: content_block_delta\ndata: {delta_event.model_dump_json()}\n\n"
-        else:
-            # Close thinking block when transitioning to text
-            if thinking_block_started and text_block_index == -1:
-                block_stop = ClaudeContentBlockStopEvent(index=thinking_block_index)
-                yield f"event: content_block_stop\ndata: {block_stop.model_dump_json()}\n\n"
-
-            # Start text block on first text token
-            if not text_block_started:
-                text_block_started = True
-                text_block_index = next_block_index
-                next_block_index += 1
-                block_start = ClaudeContentBlockStartEvent(
-                    index=text_block_index,
-                    content_block=ClaudeTextBlock(text=""),
-                )
-                yield f"event: content_block_start\ndata: {block_start.model_dump_json()}\n\n"
-
-            delta_event = ClaudeContentBlockDeltaEvent(
-                index=text_block_index,
-                delta=ClaudeTextDelta(text=chunk.text),
-            )
-            yield f"event: content_block_delta\ndata: {delta_event.model_dump_json()}\n\n"
+        # content_block_delta
+        delta_event = ClaudeContentBlockDeltaEvent(
+            index=0,
+            delta=ClaudeTextDelta(text=chunk.text),
+        )
+        yield f"event: content_block_delta\ndata: {delta_event.model_dump_json()}\n\n"

        if chunk.finish_reason is not None:
            stop_reason = finish_reason_to_claude_stop_reason(chunk.finish_reason)
@@ -402,22 +309,9 @@ async def generate_claude_stream(
    if last_usage is not None:
        output_tokens = last_usage.completion_tokens

-    # Close any open blocks
-    if thinking_block_started and text_block_index == -1:
-        block_stop = ClaudeContentBlockStopEvent(index=thinking_block_index)
-        yield f"event: content_block_stop\ndata: {block_stop.model_dump_json()}\n\n"
-
-    if text_block_started:
-        block_stop = ClaudeContentBlockStopEvent(index=text_block_index)
-        yield f"event: content_block_stop\ndata: {block_stop.model_dump_json()}\n\n"
-
-    if not thinking_block_started and not text_block_started:
-        empty_start = ClaudeContentBlockStartEvent(
-            index=0, content_block=ClaudeTextBlock(text="")
-        )
-        yield f"event: content_block_start\ndata: {empty_start.model_dump_json()}\n\n"
-        empty_stop = ClaudeContentBlockStopEvent(index=0)
-        yield f"event: content_block_stop\ndata: {empty_stop.model_dump_json()}\n\n"
+    # content_block_stop for text block
+    block_stop = ClaudeContentBlockStopEvent(index=0)
+    yield f"event: content_block_stop\ndata: {block_stop.model_dump_json()}\n\n"

    # message_delta
    message_delta = ClaudeMessageDeltaEvent(
@@ -1,456 +0,0 @@
-from __future__ import annotations
-
-import json
-from collections.abc import AsyncGenerator
-from typing import Any
-
-from exo.shared.types.chunks import (
-    ErrorChunk,
-    PrefillProgressChunk,
-    TokenChunk,
-    ToolCallChunk,
-)
-from exo.shared.types.common import CommandId
-from exo.shared.types.ollama_api import (
-    OllamaChatRequest,
-    OllamaChatResponse,
-    OllamaDoneReason,
-    OllamaGenerateRequest,
-    OllamaGenerateResponse,
-    OllamaMessage,
-    OllamaToolCall,
-    OllamaToolFunction,
-)
-from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
-
-
-def _map_done_reason(
-    finish_reason: str | None,
-) -> OllamaDoneReason | None:
-    if finish_reason is None:
-        return None
-    if finish_reason == "stop":
-        return "stop"
-    if finish_reason == "length":
-        return "length"
-    if finish_reason in ("tool_calls", "function_call"):
-        return "tool_call"
-    if finish_reason == "error":
-        return "error"
-    return "stop"
-
-
-def _try_parse_json(value: str) -> dict[str, Any] | str:
-    try:
-        return json.loads(value)  # type: ignore
-    except json.JSONDecodeError:
-        return value
-
-
-def _build_tool_calls(chunk: ToolCallChunk) -> list[OllamaToolCall]:
-    tool_calls: list[OllamaToolCall] = []
-    for index, tool in enumerate(chunk.tool_calls):
-        # tool.arguments is always str; try to parse as JSON dict for Ollama format
-        arguments: dict[str, Any] | str = _try_parse_json(tool.arguments)
-        tool_calls.append(
-            OllamaToolCall(
-                id=tool.id,
-                type="function",
-                function=OllamaToolFunction(
-                    name=tool.name, arguments=arguments, index=index
-                ),
-            )
-        )
-    return tool_calls
-
-
-def _get_usage(
-    chunk: TokenChunk | ToolCallChunk,
-) -> tuple[int | None, int | None]:
-    """Extract (prompt_eval_count, eval_count) from a chunk."""
-    if chunk.usage is not None:
-        return (chunk.usage.prompt_tokens, chunk.usage.completion_tokens)
-    if chunk.stats is not None:
-        return (chunk.stats.prompt_tokens, chunk.stats.generation_tokens)
-    return (None, None)
-
-
-def ollama_request_to_text_generation(
-    request: OllamaChatRequest,
-) -> TextGenerationTaskParams:
-    """Convert Ollama chat request to exo's internal text generation format."""
-    instructions: str | None = None
-    input_messages: list[InputMessage] = []
-    chat_template_messages: list[dict[str, Any]] = []
-    tool_message_index = 0
-
-    for msg in request.messages:
-        content = msg.content or ""
-
-        if msg.role == "system":
-            if instructions is None:
-                instructions = content
-            else:
-                instructions = f"{instructions}\n{content}"
-            chat_template_messages.append({"role": "system", "content": content})
-            continue
-
-        if msg.role in ("user", "assistant") and (
-            msg.content is not None or msg.thinking is not None or msg.tool_calls
-        ):
-            input_messages.append(InputMessage(role=msg.role, content=content))
-
-        dumped: dict[str, Any] = {"role": msg.role, "content": content}
-        if msg.thinking is not None:
-            dumped["thinking"] = msg.thinking
-        if msg.tool_calls is not None:
-            tool_calls_list: list[dict[str, Any]] = []
-            for tc in msg.tool_calls:
-                function: dict[str, Any] = {
-                    "name": tc.function.name,
-                    "arguments": (
-                        json.dumps(tc.function.arguments)
-                        if isinstance(tc.function.arguments, dict)
-                        else tc.function.arguments
-                    ),
-                }
-                if tc.function.index is not None:
-                    function["index"] = tc.function.index
-                tool_call: dict[str, Any] = {"function": function}
-                if tc.id is not None:
-                    tool_call["id"] = tc.id
-                if tc.type is not None:
-                    tool_call["type"] = tc.type
-                tool_calls_list.append(tool_call)
-            dumped["tool_calls"] = tool_calls_list
-        if msg.name is not None:
-            dumped["name"] = msg.name
-        if msg.role == "tool":
-            tool_message_index += 1
-            tool_call_id = msg.tool_name or msg.name or f"tool_{tool_message_index}"
-            dumped["tool_call_id"] = tool_call_id
-            if msg.tool_name is not None:
-                dumped["tool_name"] = msg.tool_name
-        chat_template_messages.append(dumped)
-
-    options = request.options
-    return TextGenerationTaskParams(
-        model=request.model,
-        input=input_messages
-        if input_messages
-        else [InputMessage(role="user", content="")],
-        instructions=instructions,
-        max_output_tokens=options.num_predict if options else None,
-        temperature=options.temperature if options else None,
-        top_p=options.top_p if options else None,
-        top_k=options.top_k if options else None,
-        stop=options.stop if options else None,
-        seed=options.seed if options else None,
-        stream=request.stream,
-        tools=request.tools,
-        enable_thinking=request.think,
-        chat_template_messages=chat_template_messages
-        if chat_template_messages
-        else None,
-    )
-
-
-async def generate_ollama_chat_stream(
-    _command_id: CommandId,
-    chunk_stream: AsyncGenerator[
-        ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
-    ],
-) -> AsyncGenerator[str, None]:
-    """Generate streaming responses in Ollama format (newline-delimited JSON)."""
-    thinking_parts: list[str] = []
-
-    async for chunk in chunk_stream:
-        match chunk:
-            case PrefillProgressChunk():
-                continue
-
-            case ErrorChunk():
-                error_response = OllamaChatResponse(
-                    model=str(chunk.model),
-                    message=OllamaMessage(
-                        role="assistant", content=chunk.error_message
-                    ),
-                    done=True,
-                    done_reason="error",
-                )
-                yield f"{error_response.model_dump_json(exclude_none=True)}\n"
-                return
-
-            case ToolCallChunk():
-                prompt_eval, eval_count = _get_usage(chunk)
-                response = OllamaChatResponse(
-                    model=str(chunk.model),
-                    message=OllamaMessage(
-                        role="assistant",
-                        content="",
-                        tool_calls=_build_tool_calls(chunk),
-                        thinking="".join(thinking_parts) if thinking_parts else None,
-                    ),
-                    done=True,
-                    done_reason="tool_call",
-                    prompt_eval_count=prompt_eval,
-                    eval_count=eval_count,
-                )
-                yield f"{response.model_dump_json(exclude_none=True)}\n"
-                return
-
-            case TokenChunk():
-                done = chunk.finish_reason is not None
-
-                if chunk.is_thinking:
-                    thinking_parts.append(chunk.text)
-                    response = OllamaChatResponse(
-                        model=str(chunk.model),
-                        message=OllamaMessage(
-                            role="assistant", content="", thinking=chunk.text
-                        ),
-                        done=False,
-                    )
-                    yield f"{response.model_dump_json(exclude_none=True)}\n"
-                elif done:
-                    prompt_eval, eval_count = _get_usage(chunk)
-                    response = OllamaChatResponse(
-                        model=str(chunk.model),
-                        message=OllamaMessage(
-                            role="assistant",
-                            content=chunk.text,
-                        ),
-                        done=True,
-                        done_reason=_map_done_reason(chunk.finish_reason),
-                        prompt_eval_count=prompt_eval,
-                        eval_count=eval_count,
-                    )
-                    yield f"{response.model_dump_json(exclude_none=True)}\n"
-                else:
-                    response = OllamaChatResponse(
-                        model=str(chunk.model),
-                        message=OllamaMessage(role="assistant", content=chunk.text),
-                        done=False,
-                    )
-                    yield f"{response.model_dump_json(exclude_none=True)}\n"
-
-                if done:
-                    return
-
-
-async def collect_ollama_chat_response(
-    _command_id: CommandId,
-    chunk_stream: AsyncGenerator[
-        ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
-    ],
-) -> AsyncGenerator[str]:
-    """Collect streaming chunks into a single non-streaming Ollama response.
-
-    Returns an AsyncGenerator[str] (single yield) for consistency with FastAPI
-    StreamingResponse cancellation handling.
-    """
-    text_parts: list[str] = []
-    thinking_parts: list[str] = []
-    tool_calls: list[OllamaToolCall] = []
-    model: str | None = None
-    finish_reason: str | None = None
-    prompt_eval_count: int | None = None
-    eval_count: int | None = None
-
-    async for chunk in chunk_stream:
-        match chunk:
-            case PrefillProgressChunk():
-                continue
-
-            case ErrorChunk():
-                raise ValueError(chunk.error_message or "Internal server error")
-
-            case TokenChunk():
-                if model is None:
-                    model = str(chunk.model)
-                if chunk.is_thinking:
-                    thinking_parts.append(chunk.text)
-                else:
-                    text_parts.append(chunk.text)
-                if chunk.finish_reason is not None:
-                    finish_reason = chunk.finish_reason
-                    prompt_eval_count, eval_count = _get_usage(chunk)
-
-            case ToolCallChunk():
-                if model is None:
-                    model = str(chunk.model)
-                tool_calls.extend(_build_tool_calls(chunk))
-                finish_reason = chunk.finish_reason
-                prompt_eval_count, eval_count = _get_usage(chunk)
-
-    combined_text = "".join(text_parts)
-    combined_thinking = "".join(thinking_parts) if thinking_parts else None
-    assert model is not None
-
-    yield OllamaChatResponse(
-        model=model,
-        message=OllamaMessage(
-            role="assistant",
-            content=combined_text,
-            thinking=combined_thinking,
-            tool_calls=tool_calls if tool_calls else None,
-        ),
-        done=True,
-        done_reason=_map_done_reason(finish_reason),
-        prompt_eval_count=prompt_eval_count,
-        eval_count=eval_count,
-    ).model_dump_json(exclude_none=True)
-    return
-
-
-# ── /api/generate ──
-
-
-def ollama_generate_request_to_text_generation(
-    request: OllamaGenerateRequest,
-) -> TextGenerationTaskParams:
-    """Convert Ollama generate request to exo's internal text generation format."""
-    chat_template_messages: list[dict[str, Any]] = []
-    if request.system:
-        chat_template_messages.append({"role": "system", "content": request.system})
-    chat_template_messages.append({"role": "user", "content": request.prompt})
-
-    options = request.options
-    return TextGenerationTaskParams(
-        model=request.model,
-        input=[InputMessage(role="user", content=request.prompt)],
-        instructions=request.system,
-        max_output_tokens=options.num_predict if options else None,
-        temperature=options.temperature if options else None,
-        top_p=options.top_p if options else None,
-        top_k=options.top_k if options else None,
-        stop=options.stop if options else None,
-        seed=options.seed if options else None,
-        stream=request.stream,
-        enable_thinking=request.think,
-        chat_template_messages=chat_template_messages
-        if chat_template_messages
-        else None,
-    )
-
-
-async def generate_ollama_generate_stream(
-    _command_id: CommandId,
-    chunk_stream: AsyncGenerator[
-        ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
-    ],
-) -> AsyncGenerator[str, None]:
-    """Generate streaming responses for /api/generate in Ollama NDJSON format."""
-    thinking_parts: list[str] = []
-
-    async for chunk in chunk_stream:
-        match chunk:
-            case PrefillProgressChunk():
-                continue
-
-            case ErrorChunk():
-                resp = OllamaGenerateResponse(
-                    model=str(chunk.model),
-                    response="",
-                    done=True,
-                    done_reason="error",
-                )
-                yield f"{resp.model_dump_json(exclude_none=True)}\n"
-                return
-
-            case ToolCallChunk():
-                # generate endpoint doesn't support tools; emit as done
-                prompt_eval, eval_count = _get_usage(chunk)
-                resp = OllamaGenerateResponse(
-                    model=str(chunk.model),
-                    response="",
-                    done=True,
-                    done_reason="stop",
-                    prompt_eval_count=prompt_eval,
-                    eval_count=eval_count,
-                )
-                yield f"{resp.model_dump_json(exclude_none=True)}\n"
-                return
-
-            case TokenChunk():
-                done = chunk.finish_reason is not None
-
-                if chunk.is_thinking:
-                    thinking_parts.append(chunk.text)
-                    resp = OllamaGenerateResponse(
-                        model=str(chunk.model),
-                        response="",
-                        thinking=chunk.text,
-                        done=False,
-                    )
-                    yield f"{resp.model_dump_json(exclude_none=True)}\n"
-                elif done:
-                    prompt_eval, eval_count = _get_usage(chunk)
-                    resp = OllamaGenerateResponse(
-                        model=str(chunk.model),
-                        response=chunk.text,
-                        done=True,
-                        done_reason=_map_done_reason(chunk.finish_reason),
-                        prompt_eval_count=prompt_eval,
-                        eval_count=eval_count,
-                    )
-                    yield f"{resp.model_dump_json(exclude_none=True)}\n"
-                else:
-                    resp = OllamaGenerateResponse(
-                        model=str(chunk.model),
-                        response=chunk.text,
-                        done=False,
-                    )
-                    yield f"{resp.model_dump_json(exclude_none=True)}\n"
-
-                if done:
-                    return
-
-
-async def collect_ollama_generate_response(
-    _command_id: CommandId,
-    chunk_stream: AsyncGenerator[
-        ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
-    ],
-) -> AsyncGenerator[str]:
-    """Collect chunks into a single non-streaming /api/generate response."""
-    text_parts: list[str] = []
-    thinking_parts: list[str] = []
-    model: str | None = None
-    finish_reason: str | None = None
-    prompt_eval_count: int | None = None
-    eval_count: int | None = None
-
-    async for chunk in chunk_stream:
-        match chunk:
-            case PrefillProgressChunk():
-                continue
-            case ErrorChunk():
-                raise ValueError(chunk.error_message or "Internal server error")
-            case TokenChunk():
-                if model is None:
-                    model = str(chunk.model)
-                if chunk.is_thinking:
-                    thinking_parts.append(chunk.text)
-                else:
-                    text_parts.append(chunk.text)
-                if chunk.finish_reason is not None:
-                    finish_reason = chunk.finish_reason
-                    prompt_eval_count, eval_count = _get_usage(chunk)
-            case ToolCallChunk():
-                if model is None:
-                    model = str(chunk.model)
-                finish_reason = chunk.finish_reason
-                prompt_eval_count, eval_count = _get_usage(chunk)
-
-    assert model is not None
-    yield OllamaGenerateResponse(
-        model=model,
-        response="".join(text_parts),
-        thinking="".join(thinking_parts) if thinking_parts else None,
-        done=True,
-        done_reason=_map_done_reason(finish_reason),
-        prompt_eval_count=prompt_eval_count,
-        eval_count=eval_count,
-    ).model_dump_json(exclude_none=True)
-    return
@@ -5,12 +5,7 @@ from itertools import count
 from typing import Any

 from exo.shared.types.api import Usage
-from exo.shared.types.chunks import (
-    ErrorChunk,
-    PrefillProgressChunk,
-    TokenChunk,
-    ToolCallChunk,
-)
+from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
 from exo.shared.types.common import CommandId
 from exo.shared.types.openai_responses import (
    FunctionCallInputItem,
@@ -29,15 +24,8 @@ from exo.shared.types.openai_responses import (
    ResponseOutputItemAddedEvent,
    ResponseOutputItemDoneEvent,
    ResponseOutputText,
-    ResponseReasoningItem,
-    ResponseReasoningSummaryPartAddedEvent,
-    ResponseReasoningSummaryPartDoneEvent,
-    ResponseReasoningSummaryText,
-    ResponseReasoningSummaryTextDeltaEvent,
-    ResponseReasoningSummaryTextDoneEvent,
    ResponsesRequest,
    ResponsesResponse,
-    ResponsesStreamEvent,
    ResponseTextDeltaEvent,
    ResponseTextDoneEvent,
    ResponseUsage,
@@ -45,11 +33,6 @@ from exo.shared.types.openai_responses import (
 from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams


-def _format_sse(event: ResponsesStreamEvent) -> str:
-    """Format a streaming event as an SSE message."""
-    return f"event: {event.type}\ndata: {event.model_dump_json()}\n\n"
-
-
 def _extract_content(content: str | list[ResponseContentPart]) -> str:
    """Extract plain text from a content field that may be a string or list of parts."""
    if isinstance(content, str):
@@ -138,26 +121,19 @@ def responses_request_to_text_generation(
 async def collect_responses_response(
    command_id: CommandId,
    model: str,
-    chunk_stream: AsyncGenerator[
-        ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
-    ],
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
 ) -> 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}"
-    reasoning_id = f"rs_{command_id}"
    accumulated_text = ""
-    thinking_parts: list[str] = []
    function_call_items: list[ResponseFunctionCallItem] = []
    last_usage: Usage | None = None
    error_message: str | None = None

    async for chunk in chunk_stream:
-        if isinstance(chunk, PrefillProgressChunk):
-            continue
-
        if isinstance(chunk, ErrorChunk):
            error_message = chunk.error_message or "Internal server error"
            break
@@ -168,18 +144,14 @@ async def collect_responses_response(
            for tool in chunk.tool_calls:
                function_call_items.append(
                    ResponseFunctionCallItem(
-                        id=tool.id,
-                        call_id=tool.id,
+                        id=f"fc_{tool.id}",
+                        call_id=f"call_{tool.id}",
                        name=tool.name,
                        arguments=tool.arguments,
                    )
                )
            continue

-        if chunk.is_thinking:
-            thinking_parts.append(chunk.text)
-            continue
-
        accumulated_text += chunk.text

    if error_message is not None:
@@ -194,21 +166,13 @@ async def collect_responses_response(
            total_tokens=last_usage.total_tokens,
        )

-    output: list[ResponseItem] = []
-    if thinking_parts:
-        output.append(
-            ResponseReasoningItem(
-                id=reasoning_id,
-                summary=[ResponseReasoningSummaryText(text="".join(thinking_parts))],
-            )
-        )
-    output.append(
+    output: list[ResponseItem] = [
        ResponseMessageItem(
            id=item_id,
            content=[ResponseOutputText(text=accumulated_text)],
            status="completed",
        )
-    )
+    ]
    output.extend(function_call_items)

    yield ResponsesResponse(
@@ -225,14 +189,11 @@ async def collect_responses_response(
 async def generate_responses_stream(
    command_id: CommandId,
    model: str,
-    chunk_stream: AsyncGenerator[
-        ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
-    ],
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
 ) -> AsyncGenerator[str, None]:
    """Generate OpenAI Responses API streaming events from TokenChunks."""
    response_id = f"resp_{command_id}"
    item_id = f"item_{command_id}"
-    reasoning_id = f"rs_{command_id}"
    seq = count(1)

    # response.created
@@ -246,30 +207,42 @@ async def generate_responses_stream(
    created_event = ResponseCreatedEvent(
        sequence_number=next(seq), response=initial_response
    )
-    yield _format_sse(created_event)
+    yield f"event: response.created\ndata: {created_event.model_dump_json()}\n\n"

    # response.in_progress
    in_progress_event = ResponseInProgressEvent(
        sequence_number=next(seq), response=initial_response
    )
-    yield _format_sse(in_progress_event)
+    yield f"event: response.in_progress\ndata: {in_progress_event.model_dump_json()}\n\n"
+
+    # response.output_item.added
+    initial_item = ResponseMessageItem(
+        id=item_id,
+        content=[ResponseOutputText(text="")],
+        status="in_progress",
+    )
+    item_added = ResponseOutputItemAddedEvent(
+        sequence_number=next(seq), output_index=0, item=initial_item
+    )
+    yield f"event: response.output_item.added\ndata: {item_added.model_dump_json()}\n\n"
+
+    # response.content_part.added
+    initial_part = ResponseOutputText(text="")
+    part_added = ResponseContentPartAddedEvent(
+        sequence_number=next(seq),
+        item_id=item_id,
+        output_index=0,
+        content_index=0,
+        part=initial_part,
+    )
+    yield f"event: response.content_part.added\ndata: {part_added.model_dump_json()}\n\n"

    accumulated_text = ""
-    accumulated_thinking = ""
    function_call_items: list[ResponseFunctionCallItem] = []
    last_usage: Usage | None = None
-    next_output_index = 0
-
-    # Track dynamic block creation
-    reasoning_started = False
-    reasoning_output_index = -1
-    message_started = False
-    message_output_index = -1
+    next_output_index = 1  # message item is at 0

    async for chunk in chunk_stream:
-        if isinstance(chunk, PrefillProgressChunk):
-            continue
-
        if isinstance(chunk, ErrorChunk):
            break

@@ -293,7 +266,7 @@ async def generate_responses_stream(
                    output_index=next_output_index,
                    item=fc_item,
                )
-                yield _format_sse(fc_added)
+                yield f"event: response.output_item.added\ndata: {fc_added.model_dump_json()}\n\n"

                # response.function_call_arguments.delta
                args_delta = ResponseFunctionCallArgumentsDeltaEvent(
@@ -302,7 +275,7 @@ async def generate_responses_stream(
                    output_index=next_output_index,
                    delta=tool.arguments,
                )
-                yield _format_sse(args_delta)
+                yield f"event: response.function_call_arguments.delta\ndata: {args_delta.model_dump_json()}\n\n"

                # response.function_call_arguments.done
                args_done = ResponseFunctionCallArgumentsDoneEvent(
@@ -312,7 +285,7 @@ async def generate_responses_stream(
                    name=tool.name,
                    arguments=tool.arguments,
                )
-                yield _format_sse(args_done)
+                yield f"event: response.function_call_arguments.done\ndata: {args_done.model_dump_json()}\n\n"

                # response.output_item.done
                fc_done_item = ResponseFunctionCallItem(
@@ -327,205 +300,44 @@ async def generate_responses_stream(
                    output_index=next_output_index,
                    item=fc_done_item,
                )
-                yield _format_sse(fc_item_done)
+                yield f"event: response.output_item.done\ndata: {fc_item_done.model_dump_json()}\n\n"

                function_call_items.append(fc_done_item)
                next_output_index += 1
            continue

-        if chunk.is_thinking:
-            # Start reasoning block on first thinking token
-            if not reasoning_started:
-                reasoning_started = True
-                reasoning_output_index = next_output_index
-                next_output_index += 1
-
-                # response.output_item.added for reasoning
-                reasoning_item = ResponseReasoningItem(
-                    id=reasoning_id,
-                    summary=[],
-                    status="in_progress",
-                )
-                rs_added = ResponseOutputItemAddedEvent(
-                    sequence_number=next(seq),
-                    output_index=reasoning_output_index,
-                    item=reasoning_item,
-                )
-                yield _format_sse(rs_added)
-
-                # response.reasoning_summary_part.added
-                part_added = ResponseReasoningSummaryPartAddedEvent(
-                    sequence_number=next(seq),
-                    item_id=reasoning_id,
-                    output_index=reasoning_output_index,
-                    summary_index=0,
-                    part=ResponseReasoningSummaryText(text=""),
-                )
-                yield _format_sse(part_added)
-
-            accumulated_thinking += chunk.text
-
-            # response.reasoning_summary_text.delta
-            rs_delta = ResponseReasoningSummaryTextDeltaEvent(
-                sequence_number=next(seq),
-                item_id=reasoning_id,
-                output_index=reasoning_output_index,
-                summary_index=0,
-                delta=chunk.text,
-            )
-            yield _format_sse(rs_delta)
-            continue
-
-        # Close reasoning block when transitioning to text
-        if reasoning_started and not message_started:
-            # response.reasoning_summary_text.done
-            rs_text_done = ResponseReasoningSummaryTextDoneEvent(
-                sequence_number=next(seq),
-                item_id=reasoning_id,
-                output_index=reasoning_output_index,
-                summary_index=0,
-                text=accumulated_thinking,
-            )
-            yield _format_sse(rs_text_done)
-
-            # response.reasoning_summary_part.done
-            rs_part_done = ResponseReasoningSummaryPartDoneEvent(
-                sequence_number=next(seq),
-                item_id=reasoning_id,
-                output_index=reasoning_output_index,
-                summary_index=0,
-                part=ResponseReasoningSummaryText(text=accumulated_thinking),
-            )
-            yield _format_sse(rs_part_done)
-
-            # response.output_item.done for reasoning
-            rs_item_done = ResponseOutputItemDoneEvent(
-                sequence_number=next(seq),
-                output_index=reasoning_output_index,
-                item=ResponseReasoningItem(
-                    id=reasoning_id,
-                    summary=[ResponseReasoningSummaryText(text=accumulated_thinking)],
-                ),
-            )
-            yield _format_sse(rs_item_done)
-
-        # Start message block on first text token
-        if not message_started:
-            message_started = True
-            message_output_index = next_output_index
-            next_output_index += 1
-
-            initial_item = ResponseMessageItem(
-                id=item_id,
-                content=[ResponseOutputText(text="")],
-                status="in_progress",
-            )
-            item_added = ResponseOutputItemAddedEvent(
-                sequence_number=next(seq),
-                output_index=message_output_index,
-                item=initial_item,
-            )
-            yield _format_sse(item_added)
-
-            initial_part = ResponseOutputText(text="")
-            part_added = ResponseContentPartAddedEvent(
-                sequence_number=next(seq),
-                item_id=item_id,
-                output_index=message_output_index,
-                content_index=0,
-                part=initial_part,
-            )
-            yield _format_sse(part_added)
-
        accumulated_text += chunk.text

        # response.output_text.delta
        delta_event = ResponseTextDeltaEvent(
            sequence_number=next(seq),
            item_id=item_id,
-            output_index=message_output_index,
+            output_index=0,
            content_index=0,
            delta=chunk.text,
        )
-        yield _format_sse(delta_event)
-
-    # Close reasoning block if it was never followed by text
-    if reasoning_started and not message_started:
-        rs_text_done = ResponseReasoningSummaryTextDoneEvent(
-            sequence_number=next(seq),
-            item_id=reasoning_id,
-            output_index=reasoning_output_index,
-            summary_index=0,
-            text=accumulated_thinking,
-        )
-        yield _format_sse(rs_text_done)
-
-        rs_part_done = ResponseReasoningSummaryPartDoneEvent(
-            sequence_number=next(seq),
-            item_id=reasoning_id,
-            output_index=reasoning_output_index,
-            summary_index=0,
-            part=ResponseReasoningSummaryText(text=accumulated_thinking),
-        )
-        yield _format_sse(rs_part_done)
-
-        rs_item_done = ResponseOutputItemDoneEvent(
-            sequence_number=next(seq),
-            output_index=reasoning_output_index,
-            item=ResponseReasoningItem(
-                id=reasoning_id,
-                summary=[ResponseReasoningSummaryText(text=accumulated_thinking)],
-            ),
-        )
-        yield _format_sse(rs_item_done)
-
-    # If no message block was started, create one now (empty text)
-    if not message_started:
-        message_output_index = next_output_index
-        next_output_index += 1
-
-        initial_item = ResponseMessageItem(
-            id=item_id,
-            content=[ResponseOutputText(text="")],
-            status="in_progress",
-        )
-        item_added = ResponseOutputItemAddedEvent(
-            sequence_number=next(seq),
-            output_index=message_output_index,
-            item=initial_item,
-        )
-        yield _format_sse(item_added)
-
-        initial_part = ResponseOutputText(text="")
-        part_added_evt = ResponseContentPartAddedEvent(
-            sequence_number=next(seq),
-            item_id=item_id,
-            output_index=message_output_index,
-            content_index=0,
-            part=initial_part,
-        )
-        yield _format_sse(part_added_evt)
+        yield f"event: response.output_text.delta\ndata: {delta_event.model_dump_json()}\n\n"

    # response.output_text.done
    text_done = ResponseTextDoneEvent(
        sequence_number=next(seq),
        item_id=item_id,
-        output_index=message_output_index,
+        output_index=0,
        content_index=0,
        text=accumulated_text,
    )
-    yield _format_sse(text_done)
+    yield f"event: response.output_text.done\ndata: {text_done.model_dump_json()}\n\n"

    # response.content_part.done
    final_part = ResponseOutputText(text=accumulated_text)
    part_done = ResponseContentPartDoneEvent(
        sequence_number=next(seq),
        item_id=item_id,
-        output_index=message_output_index,
+        output_index=0,
        content_index=0,
        part=final_part,
    )
-    yield _format_sse(part_done)
+    yield f"event: response.content_part.done\ndata: {part_done.model_dump_json()}\n\n"

    # response.output_item.done
    final_message_item = ResponseMessageItem(
@@ -534,11 +346,9 @@ async def generate_responses_stream(
        status="completed",
    )
    item_done = ResponseOutputItemDoneEvent(
-        sequence_number=next(seq),
-        output_index=message_output_index,
-        item=final_message_item,
+        sequence_number=next(seq), output_index=0, item=final_message_item
    )
-    yield _format_sse(item_done)
+    yield f"event: response.output_item.done\ndata: {item_done.model_dump_json()}\n\n"

    # Create usage from usage data if available
    usage = None
@@ -550,15 +360,7 @@ async def generate_responses_stream(
        )

    # response.completed
-    output: list[ResponseItem] = []
-    if reasoning_started:
-        output.append(
-            ResponseReasoningItem(
-                id=reasoning_id,
-                summary=[ResponseReasoningSummaryText(text=accumulated_thinking)],
-            )
-        )
-    output.append(final_message_item)
+    output: list[ResponseItem] = [final_message_item]
    output.extend(function_call_items)
    final_response = ResponsesResponse(
        id=response_id,
@@ -571,4 +373,4 @@ async def generate_responses_stream(
    completed_event = ResponseCompletedEvent(
        sequence_number=next(seq), response=final_response
    )
-    yield _format_sse(completed_event)
+    yield f"event: response.completed\ndata: {completed_event.model_dump_json()}\n\n"
@@ -32,14 +32,6 @@ from exo.master.adapters.claude import (
    collect_claude_response,
    generate_claude_stream,
 )
-from exo.master.adapters.ollama import (
-    collect_ollama_chat_response,
-    collect_ollama_generate_response,
-    generate_ollama_chat_stream,
-    generate_ollama_generate_stream,
-    ollama_generate_request_to_text_generation,
-    ollama_request_to_text_generation,
-)
 from exo.master.adapters.responses import (
    collect_responses_response,
    generate_responses_stream,
@@ -79,8 +71,11 @@ from exo.shared.types.api import (
    ChatCompletionResponse,
    CreateInstanceParams,
    CreateInstanceResponse,
+    CreateMetaInstanceParams,
+    CreateMetaInstanceResponse,
    DeleteDownloadResponse,
    DeleteInstanceResponse,
+    DeleteMetaInstanceResponse,
    ErrorInfo,
    ErrorResponse,
    FinishReason,
@@ -93,7 +88,6 @@ from exo.shared.types.api import (
    ImageGenerationTaskParams,
    ImageListItem,
    ImageListResponse,
-    ImageSize,
    ModelList,
    ModelListModel,
    PlaceInstanceParams,
@@ -109,13 +103,11 @@ from exo.shared.types.api import (
    TraceRankStats,
    TraceResponse,
    TraceStatsResponse,
-    normalize_image_size,
 )
 from exo.shared.types.chunks import (
    ErrorChunk,
    ImageChunk,
    InputImageChunk,
-    PrefillProgressChunk,
    TokenChunk,
    ToolCallChunk,
 )
@@ -126,8 +118,10 @@ from exo.shared.types.claude_api import (
 from exo.shared.types.commands import (
    Command,
    CreateInstance,
+    CreateMetaInstance,
    DeleteDownload,
    DeleteInstance,
+    DeleteMetaInstance,
    DownloadCommand,
    ForwarderCommand,
    ForwarderDownloadCommand,
@@ -140,34 +134,21 @@ from exo.shared.types.commands import (
    TaskFinished,
    TextGeneration,
 )
-from exo.shared.types.common import CommandId, Id, NodeId, SessionId, SystemId
+from exo.shared.types.common import CommandId, Id, MetaInstanceId, NodeId, SessionId
 from exo.shared.types.events import (
    ChunkGenerated,
    Event,
-    GlobalForwarderEvent,
+    ForwarderEvent,
    IndexedEvent,
    TracesMerged,
 )
 from exo.shared.types.memory import Memory
-from exo.shared.types.ollama_api import (
-    OllamaChatRequest,
-    OllamaChatResponse,
-    OllamaGenerateRequest,
-    OllamaGenerateResponse,
-    OllamaModelDetails,
-    OllamaModelTag,
-    OllamaPsModel,
-    OllamaPsResponse,
-    OllamaShowRequest,
-    OllamaShowResponse,
-    OllamaTagsResponse,
-)
+from exo.shared.types.meta_instance import MetaInstance
 from exo.shared.types.openai_responses import (
    ResponsesRequest,
    ResponsesResponse,
 )
 from exo.shared.types.state import State
-from exo.shared.types.worker.downloads import DownloadCompleted
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding
 from exo.utils.banner import print_startup_banner
@@ -197,7 +178,8 @@ class API:
        session_id: SessionId,
        *,
        port: int,
-        global_event_receiver: Receiver[GlobalForwarderEvent],
+        # Ideally this would be a MasterForwarderEvent but type system says no :(
+        global_event_receiver: Receiver[ForwarderEvent],
        command_sender: Sender[ForwarderCommand],
        download_command_sender: Sender[ForwarderDownloadCommand],
        # This lets us pause the API if an election is running
@@ -205,7 +187,6 @@ class API:
    ) -> None:
        self.state = State()
        self._event_log = DiskEventLog(_API_EVENT_LOG_DIR)
-        self._system_id = SystemId()
        self.command_sender = command_sender
        self.download_command_sender = download_command_sender
        self.global_event_receiver = global_event_receiver
@@ -243,8 +224,7 @@ class API:
        )

        self._text_generation_queues: dict[
-            CommandId,
-            Sender[TokenChunk | ErrorChunk | ToolCallChunk | PrefillProgressChunk],
+            CommandId, Sender[TokenChunk | ErrorChunk | ToolCallChunk]
        ] = {}
        self._image_generation_queues: dict[
            CommandId, Sender[ImageChunk | ErrorChunk]
@@ -257,7 +237,6 @@ class API:
        self._event_log.close()
        self._event_log = DiskEventLog(_API_EVENT_LOG_DIR)
        self.state = State()
-        self._system_id = SystemId()
        self.session_id = new_session_id
        self.event_buffer = OrderedBuffer[Event]()
        self._text_generation_queues = {}
@@ -303,6 +282,9 @@ class API:
        self.app.get("/instance/previews")(self.get_placement_previews)
        self.app.get("/instance/{instance_id}")(self.get_instance)
        self.app.delete("/instance/{instance_id}")(self.delete_instance)
+        self.app.get("/meta_instances")(self.list_meta_instances)
+        self.app.post("/meta_instance")(self.create_meta_instance)
+        self.app.delete("/meta_instance/{meta_instance_id}")(self.delete_meta_instance)
        self.app.get("/models")(self.get_models)
        self.app.get("/v1/models")(self.get_models)
        self.app.post("/models/add")(self.add_custom_model)
@@ -322,21 +304,6 @@ class API:
        self.app.get("/images/{image_id}")(self.get_image)
        self.app.post("/v1/messages", response_model=None)(self.claude_messages)
        self.app.post("/v1/responses", response_model=None)(self.openai_responses)
-
-        # Ollama API
-        self.app.head("/ollama/")(self.ollama_version)
-        self.app.head("/ollama/api/version")(self.ollama_version)
-        self.app.post("/ollama/api/chat", response_model=None)(self.ollama_chat)
-        self.app.post("/ollama/api/api/chat", response_model=None)(self.ollama_chat)
-        self.app.post("/ollama/api/v1/chat", response_model=None)(self.ollama_chat)
-        self.app.post("/ollama/api/generate", response_model=None)(self.ollama_generate)
-        self.app.get("/ollama/api/tags")(self.ollama_tags)
-        self.app.get("/ollama/api/api/tags")(self.ollama_tags)
-        self.app.get("/ollama/api/v1/tags")(self.ollama_tags)
-        self.app.post("/ollama/api/show")(self.ollama_show)
-        self.app.get("/ollama/api/ps")(self.ollama_ps)
-        self.app.get("/ollama/api/version")(self.ollama_version)
-
        self.app.get("/state")(lambda: self.state)
        self.app.get("/events")(self.stream_events)
        self.app.post("/download/start")(self.start_download)
@@ -347,12 +314,27 @@ class API:
        self.app.get("/v1/traces/{task_id}/raw")(self.get_trace_raw)

    async def place_instance(self, payload: PlaceInstanceParams):
+        model_card = await ModelCard.load(payload.model_id)
        command = PlaceInstance(
-            model_card=await ModelCard.load(payload.model_id),
+            model_card=model_card,
            sharding=payload.sharding,
            instance_meta=payload.instance_meta,
            min_nodes=payload.min_nodes,
        )
+
+        # Validate placement before sending — fail fast with a clear error
+        # instead of silently dropping the command in the master.
+        try:
+            get_instance_placements(
+                command,
+                topology=self.state.topology,
+                current_instances=self.state.instances,
+                node_memory=self.state.node_memory,
+                node_network=self.state.node_network,
+            )
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
        await self._send(command)

        return CreateInstanceResponse(
@@ -564,25 +546,59 @@ class API:
            instance_id=instance_id,
        )

+    def list_meta_instances(self) -> dict[MetaInstanceId, MetaInstance]:
+        return dict(self.state.meta_instances)
+
+    async def create_meta_instance(
+        self, payload: CreateMetaInstanceParams
+    ) -> CreateMetaInstanceResponse:
+        meta_instance = MetaInstance(
+            model_id=payload.model_id,
+            sharding=payload.sharding,
+            instance_meta=payload.instance_meta,
+            min_nodes=payload.min_nodes,
+            node_ids=payload.node_ids,
+        )
+        command = CreateMetaInstance(meta_instance=meta_instance)
+        await self._send(command)
+        return CreateMetaInstanceResponse(
+            message="Command received.",
+            command_id=command.command_id,
+            meta_instance_id=meta_instance.meta_instance_id,
+        )
+
+    async def delete_meta_instance(
+        self, meta_instance_id: MetaInstanceId
+    ) -> DeleteMetaInstanceResponse:
+        meta = self.state.meta_instances.get(meta_instance_id)
+        if not meta:
+            raise HTTPException(status_code=404, detail="MetaInstance not found")
+
+        # Command processor handles cascade-deleting backing instances
+        command = DeleteMetaInstance(meta_instance_id=meta_instance_id)
+        await self._send(command)
+
+        return DeleteMetaInstanceResponse(
+            message="Command received.",
+            command_id=command.command_id,
+            meta_instance_id=meta_instance_id,
+        )
+
    async def _token_chunk_stream(
        self, command_id: CommandId
-    ) -> AsyncGenerator[
-        TokenChunk | ErrorChunk | ToolCallChunk | PrefillProgressChunk, None
-    ]:
+    ) -> AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None]:
        """Yield chunks for a given command until completion.

        This is the internal low-level stream used by all API adapters.
        """
        try:
            self._text_generation_queues[command_id], recv = channel[
-                TokenChunk | ErrorChunk | ToolCallChunk | PrefillProgressChunk
+                ErrorChunk | ToolCallChunk | TokenChunk
            ]()

            with recv as token_chunks:
                async for chunk in token_chunks:
                    yield chunk
-                    if isinstance(chunk, PrefillProgressChunk):
-                        continue
                    if chunk.finish_reason is not None:
                        break

@@ -590,7 +606,7 @@ class API:
            command = TaskCancelled(cancelled_command_id=command_id)
            with anyio.CancelScope(shield=True):
                await self.command_sender.send(
-                    ForwarderCommand(origin=self._system_id, command=command)
+                    ForwarderCommand(origin=self.node_id, command=command)
                )
            raise
        finally:
@@ -609,9 +625,6 @@ class API:
        stats: GenerationStats | None = None

        async for chunk in self._token_chunk_stream(command_id):
-            if isinstance(chunk, PrefillProgressChunk):
-                continue
-
            if chunk.finish_reason == "error":
                raise HTTPException(
                    status_code=500,
@@ -800,11 +813,9 @@ class API:
        When stream=True and partial_images > 0, returns a StreamingResponse
        with SSE-formatted events for partial and final images.
        """
+        payload.model = await self._validate_image_model(ModelId(payload.model))
        payload = payload.model_copy(
-            update={
-                "model": await self._validate_image_model(ModelId(payload.model)),
-                "advanced_params": _ensure_seed(payload.advanced_params),
-            }
+            update={"advanced_params": _ensure_seed(payload.advanced_params)}
        )

        command = ImageGeneration(
@@ -938,7 +949,7 @@ class API:
            command = TaskCancelled(cancelled_command_id=command_id)
            with anyio.CancelScope(shield=True):
                await self.command_sender.send(
-                    ForwarderCommand(origin=self._system_id, command=command)
+                    ForwarderCommand(origin=self.node_id, command=command)
                )
            raise
        finally:
@@ -1024,7 +1035,7 @@ class API:
            command = TaskCancelled(cancelled_command_id=command_id)
            with anyio.CancelScope(shield=True):
                await self.command_sender.send(
-                    ForwarderCommand(origin=self._system_id, command=command)
+                    ForwarderCommand(origin=self.node_id, command=command)
                )
            raise
        finally:
@@ -1060,13 +1071,12 @@ class API:
    async def bench_image_generations(
        self, request: Request, payload: BenchImageGenerationTaskParams
    ) -> BenchImageGenerationResponse:
+        payload.model = await self._validate_image_model(ModelId(payload.model))
+
+        payload.stream = False
+        payload.partial_images = 0
        payload = payload.model_copy(
-            update={
-                "model": await self._validate_image_model(ModelId(payload.model)),
-                "stream": False,
-                "partial_images": 0,
-                "advanced_params": _ensure_seed(payload.advanced_params),
-            }
+            update={"advanced_params": _ensure_seed(payload.advanced_params)}
        )

        command = ImageGeneration(
@@ -1087,7 +1097,7 @@ class API:
        prompt: str,
        model: ModelId,
        n: int,
-        size: ImageSize,
+        size: str,
        response_format: Literal["url", "b64_json"],
        input_fidelity: Literal["low", "high"],
        stream: bool,
@@ -1157,7 +1167,7 @@ class API:
        prompt: str = Form(...),
        model: str = Form(...),
        n: int = Form(1),
-        size: str | None = Form(None),
+        size: str = Form("1024x1024"),
        response_format: Literal["url", "b64_json"] = Form("b64_json"),
        input_fidelity: Literal["low", "high"] = Form("low"),
        stream: str = Form("false"),
@@ -1183,7 +1193,7 @@ class API:
            prompt=prompt,
            model=ModelId(model),
            n=n,
-            size=normalize_image_size(size),
+            size=size,
            response_format=response_format,
            input_fidelity=input_fidelity,
            stream=stream_bool,
@@ -1219,7 +1229,7 @@ class API:
        prompt: str = Form(...),
        model: str = Form(...),
        n: int = Form(1),
-        size: str | None = Form(None),
+        size: str = Form("1024x1024"),
        response_format: Literal["url", "b64_json"] = Form("b64_json"),
        input_fidelity: Literal["low", "high"] = Form("low"),
        quality: Literal["high", "medium", "low"] = Form("medium"),
@@ -1239,7 +1249,7 @@ class API:
            prompt=prompt,
            model=ModelId(model),
            n=n,
-            size=normalize_image_size(size),
+            size=size,
            response_format=response_format,
            input_fidelity=input_fidelity,
            stream=False,
@@ -1330,163 +1340,6 @@ class API:
                media_type="application/json",
            )

-    async def _ollama_root(self) -> JSONResponse:
-        """Respond to HEAD / from Ollama CLI connectivity checks."""
-        return JSONResponse(content="Ollama is running")
-
-    async def ollama_chat(
-        self, request: Request
-    ) -> OllamaChatResponse | StreamingResponse:
-        """Ollama Chat API — accepts JSON regardless of Content-Type."""
-        body = await request.body()
-        payload = OllamaChatRequest.model_validate_json(body)
-        task_params = ollama_request_to_text_generation(payload)
-        resolved_model = await self._resolve_and_validate_text_model(
-            ModelId(task_params.model)
-        )
-        task_params = task_params.model_copy(update={"model": resolved_model})
-
-        command = TextGeneration(task_params=task_params)
-        await self._send(command)
-
-        if payload.stream:
-            return StreamingResponse(
-                generate_ollama_chat_stream(
-                    command.command_id,
-                    self._token_chunk_stream(command.command_id),
-                ),
-                media_type="application/x-ndjson",
-                headers={
-                    "Cache-Control": "no-cache",
-                    "Connection": "close",
-                    "X-Accel-Buffering": "no",
-                },
-            )
-        else:
-            return StreamingResponse(
-                collect_ollama_chat_response(
-                    command.command_id,
-                    self._token_chunk_stream(command.command_id),
-                ),
-                media_type="application/json",
-            )
-
-    async def ollama_generate(
-        self, request: Request
-    ) -> OllamaGenerateResponse | StreamingResponse:
-        """Ollama Generate API — accepts JSON regardless of Content-Type."""
-        body = await request.body()
-        payload = OllamaGenerateRequest.model_validate_json(body)
-        task_params = ollama_generate_request_to_text_generation(payload)
-        resolved_model = await self._resolve_and_validate_text_model(
-            ModelId(task_params.model)
-        )
-        task_params = task_params.model_copy(update={"model": resolved_model})
-
-        command = TextGeneration(task_params=task_params)
-        await self._send(command)
-
-        if payload.stream:
-            return StreamingResponse(
-                generate_ollama_generate_stream(
-                    command.command_id,
-                    self._token_chunk_stream(command.command_id),
-                ),
-                media_type="application/x-ndjson",
-                headers={
-                    "Cache-Control": "no-cache",
-                    "Connection": "close",
-                    "X-Accel-Buffering": "no",
-                },
-            )
-        else:
-            return StreamingResponse(
-                collect_ollama_generate_response(
-                    command.command_id,
-                    self._token_chunk_stream(command.command_id),
-                ),
-                media_type="application/json",
-            )
-
-    async def ollama_tags(self) -> OllamaTagsResponse:
-        """Returns list of models in Ollama tags format. We return the downloaded ones only."""
-
-        def none_if_empty(value: str) -> str | None:
-            return value or None
-
-        downloaded_model_ids: set[str] = set()
-        for node_downloads in self.state.downloads.values():
-            for dl in node_downloads:
-                if isinstance(dl, DownloadCompleted):
-                    downloaded_model_ids.add(dl.shard_metadata.model_card.model_id)
-
-        cards = [
-            c for c in await get_model_cards() if c.model_id in downloaded_model_ids
-        ]
-
-        now = time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
-        return OllamaTagsResponse(
-            models=[
-                OllamaModelTag(
-                    name=str(card.model_id),
-                    model=str(card.model_id),
-                    modified_at=now,
-                    size=card.storage_size.in_bytes,
-                    digest="sha256:000000000000",
-                    details=OllamaModelDetails(
-                        family=none_if_empty(card.family),
-                        quantization_level=none_if_empty(card.quantization),
-                    ),
-                )
-                for card in cards
-            ]
-        )
-
-    async def ollama_show(self, request: Request) -> OllamaShowResponse:
-        """Returns model information in Ollama show format."""
-        body = await request.body()
-        payload = OllamaShowRequest.model_validate_json(body)
-        model_name = payload.name or payload.model
-        if not model_name:
-            raise HTTPException(status_code=400, detail="name or model is required")
-        try:
-            card = await ModelCard.load(ModelId(model_name))
-        except Exception as exc:
-            raise HTTPException(
-                status_code=404, detail=f"Model not found: {model_name}"
-            ) from exc
-
-        return OllamaShowResponse(
-            modelfile=f"FROM {card.model_id}",
-            template="{{ .Prompt }}",
-            details=OllamaModelDetails(
-                family=card.family or None,
-                quantization_level=card.quantization or None,
-            ),
-        )
-
-    async def ollama_ps(self) -> OllamaPsResponse:
-        """Returns list of running models (active instances)."""
-        models: list[OllamaPsModel] = []
-        seen: set[str] = set()
-        for instance in self.state.instances.values():
-            model_id = str(instance.shard_assignments.model_id)
-            if model_id in seen:
-                continue
-            seen.add(model_id)
-            models.append(
-                OllamaPsModel(
-                    name=model_id,
-                    model=model_id,
-                    size=0,
-                )
-            )
-        return OllamaPsResponse(models=models)
-
-    async def ollama_version(self) -> dict[str, str]:
-        """Returns version information for Ollama API compatibility."""
-        return {"version": "exo v1.0"}
-
    def _calculate_total_available_memory(self) -> Memory:
        """Calculate total available memory across all nodes in bytes."""
        total_available = Memory()
@@ -1496,18 +1349,8 @@ class API:

        return total_available

-    async def get_models(self, status: str | None = Query(default=None)) -> ModelList:
-        """Returns list of available models, optionally filtered by being downloaded."""
-        cards = await get_model_cards()
-
-        if status == "downloaded":
-            downloaded_model_ids: set[str] = set()
-            for node_downloads in self.state.downloads.values():
-                for dl in node_downloads:
-                    if isinstance(dl, DownloadCompleted):
-                        downloaded_model_ids.add(dl.shard_metadata.model_card.model_id)
-            cards = [c for c in cards if c.model_id in downloaded_model_ids]
-
+    async def get_models(self) -> ModelList:
+        """Returns list of available models."""
        return ModelList(
            data=[
                ModelListModel(
@@ -1516,7 +1359,7 @@ class API:
                    name=card.model_id.short(),
                    description="",
                    tags=[],
-                    storage_size_megabytes=card.storage_size.in_mb,
+                    storage_size_megabytes=int(card.storage_size.in_mb),
                    supports_tensor=card.supports_tensor,
                    tasks=[task.value for task in card.tasks],
                    is_custom=is_custom_card(card.model_id),
@@ -1525,7 +1368,7 @@ class API:
                    base_model=card.base_model,
                    capabilities=card.capabilities,
                )
-                for card in cards
+                for card in await get_model_cards()
            ]
        )

@@ -1622,8 +1465,6 @@ class API:
    async def _apply_state(self):
        with self.global_event_receiver as events:
            async for f_event in events:
-                if f_event.session != self.session_id:
-                    continue
                if f_event.origin != self.session_id.master_node_id:
                    continue
                self.event_buffer.ingest(f_event.origin_idx, f_event.event)
@@ -1650,6 +1491,7 @@ class API:
                                await queue.send(event.chunk)
                            except BrokenResourceError:
                                self._text_generation_queues.pop(event.command_id, None)
+
                    if isinstance(event, TracesMerged):
                        self._save_merged_trace(event)

@@ -1687,12 +1529,12 @@ class API:
        while self.paused:
            await self.paused_ev.wait()
        await self.command_sender.send(
-            ForwarderCommand(origin=self._system_id, command=command)
+            ForwarderCommand(origin=self.node_id, command=command)
        )

    async def _send_download(self, command: DownloadCommand):
        await self.download_command_sender.send(
-            ForwarderDownloadCommand(origin=self._system_id, command=command)
+            ForwarderDownloadCommand(origin=self.node_id, command=command)
        )

    async def start_download(
@@ -1,4 +1,5 @@
-from datetime import datetime, timedelta, timezone
+from collections.abc import Sequence
+from datetime import datetime, timezone

 import anyio
 from anyio.abc import TaskGroup
@@ -12,11 +13,22 @@ from exo.master.placement import (
    get_transition_events,
    place_instance,
 )
+from exo.master.process_managers import ProcessManager
+from exo.master.process_managers.instance_health import InstanceHealthReconciler
+from exo.master.process_managers.meta_instance import MetaInstanceReconciler
+from exo.master.process_managers.node_timeout import NodeTimeoutReconciler
+from exo.master.reconcile import (
+    find_unsatisfied_meta_instances,
+    try_place_for_meta_instance,
+)
 from exo.shared.apply import apply
 from exo.shared.constants import EXO_EVENT_LOG_DIR, EXO_TRACING_ENABLED
+from exo.shared.models.model_cards import ModelCard
 from exo.shared.types.commands import (
    CreateInstance,
+    CreateMetaInstance,
    DeleteInstance,
+    DeleteMetaInstance,
    ForwarderCommand,
    ForwarderDownloadCommand,
    ImageEdits,
@@ -29,18 +41,19 @@ from exo.shared.types.commands import (
    TestCommand,
    TextGeneration,
 )
-from exo.shared.types.common import CommandId, NodeId, SessionId, SystemId
+from exo.shared.types.common import CommandId, NodeId, SessionId
 from exo.shared.types.events import (
    Event,
-    GlobalForwarderEvent,
+    ForwarderEvent,
    IndexedEvent,
    InputChunkReceived,
    InstanceDeleted,
    JacclSideChannelData,
    JacclSideChannelGathered,
-    LocalForwarderEvent,
+    MetaInstanceCreated,
+    MetaInstanceDeleted,
+    MetaInstancePlacementFailed,
    NodeGatheredInfo,
-    NodeTimedOut,
    TaskCreated,
    TaskDeleted,
    TaskStatusUpdated,
@@ -64,7 +77,7 @@ from exo.shared.types.tasks import (
 )
 from exo.shared.types.worker.instances import InstanceId
 from exo.shared.types.worker.runners import RunnerId
-from exo.utils.channels import Receiver, Sender, channel
+from exo.utils.channels import Receiver, Sender
 from exo.utils.event_buffer import MultiSourceBuffer


@@ -75,8 +88,8 @@ class Master:
        session_id: SessionId,
        *,
        command_receiver: Receiver[ForwarderCommand],
-        local_event_receiver: Receiver[LocalForwarderEvent],
-        global_event_sender: Sender[GlobalForwarderEvent],
+        local_event_receiver: Receiver[ForwarderEvent],
+        global_event_sender: Sender[ForwarderEvent],
        download_command_sender: Sender[ForwarderDownloadCommand],
    ):
        self.state = State()
@@ -88,18 +101,16 @@ class Master:
        self.local_event_receiver = local_event_receiver
        self.global_event_sender = global_event_sender
        self.download_command_sender = download_command_sender
-        send, recv = channel[Event]()
-        self.event_sender: Sender[Event] = send
-        self._loopback_event_receiver: Receiver[Event] = recv
-        self._loopback_event_sender: Sender[LocalForwarderEvent] = (
-            local_event_receiver.clone_sender()
-        )
-        self._system_id = SystemId()
-        self._multi_buffer = MultiSourceBuffer[SystemId, Event]()
+        self._multi_buffer = MultiSourceBuffer[NodeId, Event]()
        self._event_log = DiskEventLog(EXO_EVENT_LOG_DIR / "master")
        self._pending_traces: dict[TaskId, dict[int, list[TraceEventData]]] = {}
        self._expected_ranks: dict[TaskId, set[int]] = {}
        self._jaccl_pending: dict[InstanceId, dict[int, dict[RunnerId, bytes]]] = {}
+        self._process_managers: Sequence[ProcessManager] = [
+            InstanceHealthReconciler(),
+            NodeTimeoutReconciler(),
+            MetaInstanceReconciler(),
+        ]

    async def run(self):
        logger.info("Starting Master")
@@ -108,15 +119,12 @@ class Master:
            async with self._tg as tg:
                tg.start_soon(self._event_processor)
                tg.start_soon(self._command_processor)
-                tg.start_soon(self._loopback_processor)
-                tg.start_soon(self._plan)
+                tg.start_soon(self._reconcile)
        finally:
            self._event_log.close()
            self.global_event_sender.close()
            self.local_event_receiver.close()
            self.command_receiver.close()
-            self._loopback_event_sender.close()
-            self._loopback_event_receiver.close()

    async def shutdown(self):
        logger.info("Stopping Master")
@@ -294,10 +302,89 @@ class Master:
                            ):
                                await self.download_command_sender.send(
                                    ForwarderDownloadCommand(
-                                        origin=self._system_id, command=cmd
+                                        origin=self.node_id, command=cmd
                                    )
                                )
                            generated_events.extend(transition_events)
+                        case CreateMetaInstance():
+                            logger.info(
+                                f"Creating MetaInstance for {command.meta_instance.model_id}"
+                                f" (min_nodes={command.meta_instance.min_nodes},"
+                                f" sharding={command.meta_instance.sharding})"
+                            )
+                            # Apply immediately so self.state is fresh across
+                            # the await below and the reconciler won't race.
+                            await self._apply_and_broadcast(
+                                MetaInstanceCreated(meta_instance=command.meta_instance)
+                            )
+                            # Immediate placement attempt for responsiveness
+                            model_card = await ModelCard.load(
+                                command.meta_instance.model_id
+                            )
+                            # Re-check: reconciler may have satisfied it during the await
+                            meta_id = command.meta_instance.meta_instance_id
+                            still_unsatisfied = any(
+                                m.meta_instance_id == meta_id
+                                for m in find_unsatisfied_meta_instances(
+                                    self.state.meta_instances,
+                                    self.state.instances,
+                                    self.state.topology,
+                                )
+                            )
+                            if still_unsatisfied:
+                                result = try_place_for_meta_instance(
+                                    command.meta_instance,
+                                    model_card,
+                                    self.state.topology,
+                                    self.state.instances,
+                                    self.state.node_memory,
+                                    self.state.node_network,
+                                )
+                                generated_events.extend(result.events)
+                                if result.error is not None:
+                                    generated_events.append(
+                                        MetaInstancePlacementFailed(
+                                            meta_instance_id=meta_id,
+                                            reason=result.error,
+                                        )
+                                    )
+                        case DeleteMetaInstance():
+                            backing_count = sum(
+                                1
+                                for inst in self.state.instances.values()
+                                if inst.meta_instance_id == command.meta_instance_id
+                            )
+                            logger.info(
+                                f"Deleting MetaInstance {command.meta_instance_id}"
+                                f" (cascade-deleting {backing_count} backing instance(s))"
+                            )
+                            generated_events.append(
+                                MetaInstanceDeleted(
+                                    meta_instance_id=command.meta_instance_id
+                                )
+                            )
+                            # Cascade-delete backing instances atomically,
+                            # cancelling any active tasks first.
+                            for iid, inst in self.state.instances.items():
+                                if inst.meta_instance_id == command.meta_instance_id:
+                                    for task in self.state.tasks.values():
+                                        if (
+                                            task.instance_id == iid
+                                            and task.task_status
+                                            in (
+                                                TaskStatus.Pending,
+                                                TaskStatus.Running,
+                                            )
+                                        ):
+                                            generated_events.append(
+                                                TaskStatusUpdated(
+                                                    task_status=TaskStatus.Cancelled,
+                                                    task_id=task.task_id,
+                                                )
+                                            )
+                                    generated_events.append(
+                                        InstanceDeleted(instance_id=iid)
+                                    )
                        case PlaceInstance():
                            placement = place_instance(
                                command,
@@ -360,31 +447,32 @@ class Master:
                            ):
                                await self._send_event(IndexedEvent(idx=i, event=event))
                    for event in generated_events:
-                        await self.event_sender.send(event)
+                        await self._apply_and_broadcast(event)
                except ValueError as e:
                    logger.opt(exception=e).warning("Error in command processor")

-    # These plan loops are the cracks showing in our event sourcing architecture - more things could be commands
-    async def _plan(self) -> None:
+    async def _apply_and_broadcast(self, event: Event) -> None:
+        """Apply event to state, persist to disk, and broadcast to workers.
+
+        State is updated synchronously (before any await), so callers can
+        rely on ``self.state`` reflecting this event immediately after the
+        call.  Python's cooperative scheduling guarantees no interleaving
+        between the state read and write.
+        """
+        logger.debug(f"Master indexing event: {str(event)[:100]}")
+        indexed = IndexedEvent(event=event, idx=len(self._event_log))
+        self.state = apply(self.state, indexed)
+        event._master_time_stamp = datetime.now(tz=timezone.utc)  # pyright: ignore[reportPrivateUsage]
+        self._event_log.append(event)
+        await self._send_event(indexed)
+
+    async def _reconcile(self) -> None:
        while True:
-            # kill broken instances
-            connected_node_ids = set(self.state.topology.list_nodes())
-            for instance_id, instance in self.state.instances.items():
-                for node_id in instance.shard_assignments.node_to_runner:
-                    if node_id not in connected_node_ids:
-                        await self.event_sender.send(
-                            InstanceDeleted(instance_id=instance_id)
-                        )
-                        break
-
-            # time out dead nodes
-            for node_id, time in self.state.last_seen.items():
-                now = datetime.now(tz=timezone.utc)
-                if now - time > timedelta(seconds=30):
-                    logger.info(f"Manually removing node {node_id} due to inactivity")
-                    await self.event_sender.send(NodeTimedOut(node_id=node_id))
-
-            await anyio.sleep(10)
+            for pm in self._process_managers:
+                events = await pm.reconcile(self.state)
+                for event in events:
+                    await self._apply_and_broadcast(event)
+            await anyio.sleep(1)

    async def _event_processor(self) -> None:
        with self.local_event_receiver as local_events:
@@ -402,43 +490,21 @@ class Master:
                        await self._handle_traces_collected(event)
                        continue

-                    logger.debug(f"Master indexing event: {str(event)[:100]}")
-                    indexed = IndexedEvent(event=event, idx=len(self._event_log))
-                    self.state = apply(self.state, indexed)
+                    if isinstance(event, JacclSideChannelData):
+                        await self._apply_and_broadcast(event)
+                        await self._handle_jaccl_side_channel(event)
+                        continue

-                    event._master_time_stamp = datetime.now(tz=timezone.utc)  # pyright: ignore[reportPrivateUsage]
                    if isinstance(event, NodeGatheredInfo):
                        event.when = str(datetime.now(tz=timezone.utc))

-                    self._event_log.append(event)
-                    await self._send_event(indexed)
-
-                    # After broadcasting JacclSideChannelData, accumulate and
-                    # emit gathered result when all runners have contributed.
-                    if isinstance(event, JacclSideChannelData):
-                        await self._handle_jaccl_side_channel(event)
-
-    async def _loopback_processor(self) -> None:
-        # this would ideally not be necessary.
-        # this is WAY less hacky than how I was working around this before
-        local_index = 0
-        with self._loopback_event_receiver as events:
-            async for event in events:
-                await self._loopback_event_sender.send(
-                    LocalForwarderEvent(
-                        origin=self._system_id,
-                        origin_idx=local_index,
-                        session=self.session_id,
-                        event=event,
-                    )
-                )
-                local_index += 1
+                    await self._apply_and_broadcast(event)

    # This function is re-entrant, take care!
    async def _send_event(self, event: IndexedEvent):
        # Convenience method since this line is ugly
        await self.global_event_sender.send(
-            GlobalForwarderEvent(
+            ForwarderEvent(
                origin=self.node_id,
                origin_idx=event.idx,
                session=self.session_id,
@@ -464,7 +530,7 @@ class Master:
        for trace_data in self._pending_traces[task_id].values():
            all_trace_data.extend(trace_data)

-        await self.event_sender.send(
+        await self._apply_and_broadcast(
            TracesMerged(task_id=task_id, traces=all_trace_data)
        )

@@ -503,7 +569,7 @@ class Master:
            if not self._jaccl_pending[iid]:
                del self._jaccl_pending[iid]

-            await self.event_sender.send(
+            await self._apply_and_broadcast(
                JacclSideChannelGathered(
                    instance_id=iid,
                    sequence=seq,
@@ -6,11 +6,11 @@ from typing import Sequence
 from exo.master.placement_utils import (
    Cycle,
    filter_cycles_by_memory,
+    get_largest_cycles,
    get_mlx_jaccl_coordinators,
    get_mlx_jaccl_devices_matrix,
    get_mlx_ring_hosts_by_node,
    get_shard_assignments,
-    get_smallest_cycles,
 )
 from exo.shared.models.model_cards import ModelId
 from exo.shared.topology import Topology
@@ -106,23 +106,27 @@ def place_instance(
            "Pipeline parallelism is not supported for DeepSeek V3.1 (8-bit)"
        )

-    smallest_cycles = get_smallest_cycles(cycles_with_sufficient_memory)
+    largest_cycles = get_largest_cycles(cycles_with_sufficient_memory)

-    smallest_rdma_cycles = [
-        cycle for cycle in smallest_cycles if topology.is_rdma_cycle(cycle)
+    largest_rdma_cycles = [
+        cycle for cycle in largest_cycles if topology.is_rdma_cycle(cycle)
    ]

-    if command.instance_meta == InstanceMeta.MlxJaccl and smallest_rdma_cycles != []:
-        smallest_cycles = smallest_rdma_cycles
+    if command.instance_meta == InstanceMeta.MlxJaccl:
+        if not largest_rdma_cycles:
+            raise ValueError(
+                "Requested RDMA (MlxJaccl) but no RDMA-connected cycles available"
+            )
+        largest_cycles = largest_rdma_cycles

    cycles_with_leaf_nodes: list[Cycle] = [
        cycle
-        for cycle in smallest_cycles
+        for cycle in largest_cycles
        if any(topology.node_is_leaf(node_id) for node_id in cycle)
    ]

    selected_cycle = max(
-        cycles_with_leaf_nodes if cycles_with_leaf_nodes != [] else smallest_cycles,
+        cycles_with_leaf_nodes if cycles_with_leaf_nodes != [] else largest_cycles,
        key=lambda cycle: sum(
            (node_memory[node_id].ram_available for node_id in cycle),
            start=Memory(),
@@ -141,29 +145,15 @@ def place_instance(
    if len(selected_cycle) == 1:
        command.instance_meta = InstanceMeta.MlxRing

+    # TODO: Single node instances
    match command.instance_meta:
        case InstanceMeta.MlxJaccl:
-            # TODO(evan): shard assignments should contain information about ranks, this is ugly
-            def get_device_rank(node_id: NodeId) -> int:
-                runner_id = shard_assignments.node_to_runner[node_id]
-                shard_metadata = shard_assignments.runner_to_shard.get(runner_id)
-                assert shard_metadata is not None
-                return shard_metadata.device_rank
-
-            zero_node_ids = [
-                node_id
-                for node_id in selected_cycle.node_ids
-                if get_device_rank(node_id) == 0
-            ]
-            assert len(zero_node_ids) == 1
-            coordinator_node_id = zero_node_ids[0]
-
            mlx_jaccl_devices = get_mlx_jaccl_devices_matrix(
                [node_id for node_id in selected_cycle],
                cycle_digraph,
            )
            mlx_jaccl_coordinators = get_mlx_jaccl_coordinators(
-                coordinator=coordinator_node_id,
+                coordinator=selected_cycle.node_ids[0],
                coordinator_port=random_ephemeral_port(),
                cycle_digraph=cycle_digraph,
                node_network=node_network,
@@ -37,11 +37,11 @@ def filter_cycles_by_memory(
    return filtered_cycles


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


 def allocate_layers_proportionally(
@@ -102,21 +102,22 @@ def _allocate_and_validate_layers(
    layer_allocations = allocate_layers_proportionally(
        total_layers=model_card.n_layers,
        memory_fractions=[
-            node_memory[node_id].ram_available / total_memory for node_id in node_ids
+            node_memory[node_id].ram_available.in_bytes / total_memory.in_bytes
+            for node_id in node_ids
        ],
    )

-    total_storage = model_card.storage_size
+    total_storage_bytes = model_card.storage_size.in_bytes
    total_layers = model_card.n_layers
    for i, node_id in enumerate(node_ids):
        node_layers = layer_allocations[i]
-        required_memory = (total_storage * node_layers) // total_layers
-        available_memory = node_memory[node_id].ram_available
+        required_memory = (total_storage_bytes * node_layers) // total_layers
+        available_memory = node_memory[node_id].ram_available.in_bytes
        if required_memory > available_memory:
            raise ValueError(
                f"Node {i} ({node_id}) has insufficient memory: "
-                f"requires {required_memory.in_gb:.2f} GB for {node_layers} layers, "
-                f"but only has {available_memory.in_gb:.2f} GB available"
+                f"requires {required_memory / (1024**3):.2f} GB for {node_layers} layers, "
+                f"but only has {available_memory / (1024**3):.2f} GB available"
            )

    return layer_allocations
@@ -341,7 +342,6 @@ def _find_ip_prioritised(
    other_node_id: NodeId,
    cycle_digraph: Topology,
    node_network: Mapping[NodeId, NodeNetworkInfo],
-    ring: bool,
 ) -> str | None:
    """Find an IP address between nodes with prioritization.

@@ -354,27 +354,13 @@ def _find_ip_prioritised(
    ip_to_type = {
        iface.ip_address: iface.interface_type for iface in other_network.interfaces
    }
-
-    # Ring should prioritise fastest connection. As a best-effort, we prioritise TB.
-    # TODO: Profile and get actual connection speeds.
-    if ring:
-        priority = {
-            "thunderbolt": 0,
-            "maybe_ethernet": 1,
-            "ethernet": 2,
-            "wifi": 3,
-            "unknown": 4,
-        }
-
-    # RDMA prefers ethernet coordinator
-    else:
-        priority = {
-            "ethernet": 0,
-            "wifi": 1,
-            "unknown": 2,
-            "maybe_ethernet": 3,
-            "thunderbolt": 4,
-        }
+    priority = {
+        "ethernet": 0,
+        "wifi": 1,
+        "unknown": 2,
+        "maybe_ethernet": 3,
+        "thunderbolt": 4,
+    }
    return min(ips, key=lambda ip: priority.get(ip_to_type.get(ip, "unknown"), 2))


@@ -414,7 +400,7 @@ def get_mlx_ring_hosts_by_node(
                continue

            connection_ip = _find_ip_prioritised(
-                node_id, other_node_id, cycle_digraph, node_network, ring=True
+                node_id, other_node_id, cycle_digraph, node_network
            )
            if connection_ip is None:
                raise ValueError(
@@ -445,9 +431,7 @@ def get_mlx_jaccl_coordinators(
        if n == coordinator:
            return "0.0.0.0"

-        ip = _find_ip_prioritised(
-            n, coordinator, cycle_digraph, node_network, ring=False
-        )
+        ip = _find_ip_prioritised(n, coordinator, cycle_digraph, node_network)
        if ip is not None:
            return ip

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