custom discovery take 2

Cargo.toml

@@ -1,6 +1,6 @@
 [workspace]
 resolver = "3"
-members = ["rust/networking", "rust/exo_pyo3_bindings", "rust/util"]
+members = ["rust/exo_net", "rust/networking"]
 
 [workspace.package]
 version = "0.0.1"
@@ -20,30 +20,33 @@ opt-level = 3
 [workspace.dependencies]
 ## Crate members as common dependencies
 networking = { path = "rust/networking" }
-util = { path = "rust/util" }
 
-# Macro dependecies
+# pyo3
+pyo3 = "0.27.2"
+pyo3-async-runtimes = "0.27.0"
+pyo3-log = "0.13.2"
+pyo3-stub-gen = "0.22.2"
+
+# util
 extend = "1.2"
-delegate = "0.13"
-
-# Utility dependencies
-keccak-const = "0.2"
-
-# Async dependencies
-async-stream = "0.3"
 tokio = "1.46"
 futures-lite = "2.6.1"
-futures-timer = "3.0"
-
-# Data structures
-either = "1.15"
+async-stream = "0.3.6"
+pin-project = "1.1.10"
+serde_json = "1.0.149"
+rand = "0.10.1"
+parking_lot = "0.12.5"
 
 # Tracing/logging
 log = "0.4"
+env_logger = "0.11.10"
 
 # networking
-libp2p = "0.56"
-libp2p-tcp = "0.44"
+zenoh = "=1.9.0"
+zenoh-plugin-storage-manager = { version = "=1.9.0", default-features = false }
+zenoh-plugin-trait = "=1.9.0"
+netwatcher = "0.6.0"
+bytemuck = "1.25.0"
 
 [workspace.lints.rust]
 static_mut_refs = "warn"      # Or use "warn" instead of deny

flake.lock

@@ -47,11 +47,11 @@
         "rust-analyzer-src": "rust-analyzer-src"
       },
       "locked": {
-        "lastModified": 1775807984,
-        "narHash": "sha256-Redoe3D9zGN5I9QPHWL9vfMVQBehY1fKsMiRXQ83X3w=",
+        "lastModified": 1777708550,
+        "narHash": "sha256-Qif3UXT0l5OQq8H9pRWt4/ia4gF48MWK2oHKL8uVx8U=",
         "owner": "nix-community",
         "repo": "fenix",
-        "rev": "fcf90c0c4d368b2ca917a7afa6d08e98a397e5fd",
+        "rev": "74c1591efaff494756b8d35ebe357c6c2bbdca96",
         "type": "github"
       },
       "original": {
@@ -218,11 +218,11 @@
     "rust-analyzer-src": {
       "flake": false,
       "locked": {
-        "lastModified": 1775745684,
-        "narHash": "sha256-8MbfLwd60FNa8dRFkjE+G3TT/x21G3Rsplm1bMBQUtU=",
+        "lastModified": 1777639980,
+        "narHash": "sha256-6d7Hdurvbjc5uwJuc0YiK7rZBGj6Gs3uzfBFcTs+xCc=",
         "owner": "rust-lang",
         "repo": "rust-analyzer",
-        "rev": "64ddb549bc9a70d011328746fa46a8883f937b6b",
+        "rev": "64cdaeb06f69b6b769a492edd88b022ae88e8ca2",
         "type": "github"
       },
       "original": {

flake.nix

@@ -110,7 +110,7 @@
               nixpkgs-fmt.enable = true;
               ruff-format = {
                 enable = true;
-                excludes = [ "rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi" ];
+                excludes = [ "rust/exo_net/exo_net.pyi" ];
               };
               rustfmt = {
                 enable = true;
@@ -146,7 +146,7 @@
                 config.treefmt.build.wrapper
 
                 # PYTHON
-                self'.packages.editableVenv
+                #self'.packages.editableVenv
                 uv
 
                 # RUST

justfile

@@ -22,8 +22,8 @@ sync-clean:
     uv sync --all-packages --force-reinstall --no-cache
 
 rust-rebuild:
-    PYO3_PYTHON="$(uv run python -c 'import sys; print(sys.executable)')" cargo run --bin stub_gen
-    uv sync --reinstall-package exo_pyo3_bindings
+    cargo run --bin stub_gen
+    uv sync --reinstall-package exo_net
 
 build-dashboard:
     #!/usr/bin/env bash

pyproject.toml

@@ -15,7 +15,7 @@ dependencies = [
   "huggingface-hub>=1.8.0",
   "psutil>=7.0.0",
   "loguru>=0.7.3",
-  "exo-pyo3-bindings",                         # rust bindings
+  "exo-net",                                   # rust bindings
   "anyio==4.11.0",
   "mlx==0.31.2; sys_platform == 'darwin'",
   "mlx-lm; sys_platform=='darwin'",
@@ -75,10 +75,10 @@ cuda13 = [
 ###
 
 [tool.uv.workspace]
-members = ["rust/exo_pyo3_bindings", "bench"]
+members = ["rust/exo_net", "bench"]
 
 [tool.uv.sources]
-exo-pyo3-bindings = { workspace = true }
+exo-net = { workspace = true }
 mlx = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git", branch = "address-rdma-gpu-locks", marker = "sys_platform == 'darwin'" }
 mlx-lm = { git = "https://github.com/rltakashige/mlx-lm", branch = "leo/deepseek-v4" }
 torch = [

python/parts.nix

@@ -35,17 +35,18 @@ let
         # Replace workspace exo_pyo3_bindings with Nix-built wheel.
         # Preserve passthru so mkVirtualEnv can resolve dependency groups.
         # Copy .pyi stub + py.typed marker so basedpyright can find the types.
-        exo-pyo3-bindings = pkgs.stdenv.mkDerivation {
-          pname = "exo-pyo3-bindings";
+        exo-net = pkgs.stdenv.mkDerivation {
+          pname = "exo-net";
           version = "0.1.0";
-          src = self'.packages.exo_pyo3_bindings;
+          src = self'.packages.exo-net;
           # Install from pre-built wheel
           nativeBuildInputs = [ final.pyprojectWheelHook ];
           dontStrip = true;
           passthru = prev.exo-pyo3-bindings.passthru or { };
+
           postInstall = ''
-            local siteDir=$out/${final.python.sitePackages}/exo_pyo3_bindings
-            cp ${inputs.self}/rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi $siteDir/
+            local siteDir=$out/${final.python.sitePackages}/exo_net
+            cp ${inputs.self}/rust/exo_net/exo_net.pyi $siteDir/
             touch $siteDir/py.typed
           '';
         };

result

@@ -0,0 +1 @@
+/nix/store/3945qvxy6hla1aa6dkzlcv99d2d4bw56-exo

rust/exo_net/Cargo.toml

@@ -0,0 +1,51 @@
+[package]
+name = "exo_net"
+version = { workspace = true }
+edition = { workspace = true }
+publish = false
+
+[lib]
+doctest = false
+path = "src/lib.rs"
+name = "exo_net"
+
+# "cdylib" needed to produce shared library for Python to import
+# "rlib" needed for stub-gen to run
+crate-type = ["cdylib", "rlib"]
+
+[[bin]]
+path = "src/bin/stub_gen.rs"
+name = "stub_gen"
+doc = false
+
+[lints]
+workspace = true
+
+[dependencies]
+networking.workspace = true
+extend.workspace = true
+
+# interop
+pyo3 = { workspace = true, features = ["experimental-async"] }
+pyo3-stub-gen.workspace = true
+pyo3-async-runtimes = { workspace = true, features = [
+  "attributes",
+  "tokio-runtime",
+  "testing",
+] }
+pyo3-log.workspace = true
+
+# async runtime
+tokio = { workspace = true, features = ["full"] }
+futures-lite.workspace = true
+pin-project.workspace = true
+
+# Tracing
+log.workspace = true
+env_logger.workspace = true
+
+# Networking
+zenoh.workspace = true
+rand.workspace = true
+serde_json.workspace = true
+parking_lot.workspace = true

rust/exo_net/exo_net.pyi

@@ -0,0 +1,76 @@
+# This file is automatically generated by pyo3_stub_gen
+# ruff: noqa: E501, F401, F403, F405
+
+import builtins
+import collections.abc
+import typing
+
+__all__ = [
+    "NetReceiver",
+    "NetSender",
+    "NetworkingHandle",
+    "PyFromSwarm",
+    "PySession",
+    "StateProxy",
+]
+
+@typing.final
+class NetReceiver:
+    def recv(self) -> collections.abc.Awaitable[bytes | None]: ...
+
+@typing.final
+class NetSender:
+    def send(self, data: bytes) -> collections.abc.Awaitable[bool]: ...
+
+@typing.final
+class NetworkingHandle:
+    @staticmethod
+    def new(identity: bytes, bootstrap_peers: typing.Sequence[builtins.str], listen_port: builtins.int) -> tuple[NetworkingHandle, PySession]: ...
+    async def gossipsub_subscribe(self, topic: builtins.str) -> builtins.bool:
+        r"""
+        Subscribe to a `GossipSub` topic.
+        
+        Returns `True` if the subscription worked. Returns `False` if we were already subscribed.
+        """
+    async def gossipsub_unsubscribe(self, topic: builtins.str) -> builtins.bool:
+        r"""
+        Unsubscribes from a `GossipSub` topic.
+        
+        Returns `True` if we were subscribed to this topic. Returns `False` if we were not subscribed.
+        """
+    async def gossipsub_publish(self, topic: builtins.str, data: bytes) -> None:
+        r"""
+        Publishes a message with multiple topics to the `GossipSub` network.
+        
+        If no peers are found that subscribe to this topic, throws `NoPeersSubscribedToTopicError` exception.
+        """
+    async def recv(self) -> PyFromSwarm: ...
+
+class PyFromSwarm:
+    @typing.final
+    class Connection(PyFromSwarm):
+        __match_args__ = ("connected",)
+        @property
+        def connected(self) -> builtins.bool: ...
+        def __new__(cls, connected: builtins.bool) -> PyFromSwarm.Connection: ...
+    
+    @typing.final
+    class Message(PyFromSwarm):
+        __match_args__ = ("topic", "data",)
+        @property
+        def topic(self) -> builtins.str: ...
+        @property
+        def data(self) -> bytes: ...
+        def __new__(cls, topic: builtins.str, data: bytes) -> PyFromSwarm.Message: ...
+    
+
+@typing.final
+class PySession:
+    def net_receiver(self, key: builtins.str) -> NetReceiver: ...
+    def net_sender(self, key: builtins.str) -> NetSender: ...
+    def state_proxy(self) -> StateProxy: ...
+
+@typing.final
+class StateProxy:
+    def snapshot(self) -> collections.abc.Awaitable[str]: ...
+

rust/exo_net/exo_pyo3_bindings.pyi

@@ -0,0 +1,24 @@
+# This file is automatically generated by pyo3_stub_gen
+# ruff: noqa: E501, F401
+
+import builtins
+import collections.abc
+import typing
+
+@typing.final
+class NetReceiver: ...
+
+@typing.final
+class NetSender: ...
+
+@typing.final
+class PySession:
+    @staticmethod
+    def init() -> collections.abc.Awaitable[PySession]: ...
+    def net_receiver(self, key: builtins.str) -> NetReceiver: ...
+    def net_sender(self, key: builtins.str) -> NetSender: ...
+    def state_proxy(self) -> StateProxy: ...
+
+@typing.final
+class StateProxy:
+    def snapshot(self) -> collections.abc.Awaitable[str]: ...

rust/exo_net/pyproject.toml

@@ -3,24 +3,22 @@ requires = ["maturin>=1.0,<2.0"]
 build-backend = "maturin"
 
 [project]
-name = "exo_pyo3_bindings"
-version = "0.2.1"
+name = "exo_net"
+version = "0.3.0"
 description = "Add your description here"
 readme = "README.md"
 authors = [
-  { name = "Andrei Cravtov", email = "the.andrei.cravtov@gmail.com" },
   { name = "Evan Quiney", email = "evanev7@gmail.com" },
+  { name = "Andrei Cravtov", email = "the.andrei.cravtov@gmail.com" },
 ]
 requires-python = ">=3.13"
 dependencies = []
 
 [dependency-groups]
-dev = ["exo_pyo3_bindings", "pytest>=8.4.0", "pytest-asyncio>=1.0.0"]
+dev = ["exo-net", "pytest>=8.4.0", "pytest-asyncio>=1.0.0"]
 
 [tool.maturin]
-#purelib = true
-#python-source = "python"
-module-name = "exo_pyo3_bindings"
+module-name = "exo_net"
 features = ["pyo3/extension-module", "pyo3/experimental-async"]
 
 [tool.pytest.ini_options]

rust/exo_net/src/bin/stub_gen.rs

@@ -2,7 +2,7 @@ use pyo3_stub_gen::Result;
 
 fn main() -> Result<()> {
     env_logger::Builder::from_env(env_logger::Env::default().filter_or("RUST_LOG", "info")).init();
-    let stub = exo_pyo3_bindings::stub_info()?;
+    let stub = exo_net::stub_info()?;
     stub.generate()?;
     Ok(())
 }

rust/exo_net/src/lib.rs

@@ -5,21 +5,21 @@
 //!
 
 mod allow_threading;
-mod ident;
+// mod ident;
 mod networking;
+mod point_to_point;
+mod session;
+mod state;
 
-use crate::ident::PyKeypair;
 use crate::networking::networking_submodule;
+use crate::point_to_point::{NetReceiver, NetSender};
+use crate::session::PySession;
+use crate::state::StateProxy;
 use pyo3::prelude::PyModule;
 use pyo3::types::PyModuleMethods;
-use pyo3::{Bound, PyResult, pyclass, pymodule};
+use pyo3::{Bound, PyResult, pymodule};
 use pyo3_stub_gen::define_stub_info_gatherer;
 
-/// Namespace for all the constants used by this crate.
-pub(crate) mod r#const {
-    pub const MPSC_CHANNEL_SIZE: usize = 1024;
-}
-
 /// Namespace for crate-wide extension traits/methods
 pub(crate) mod ext {
     use crate::allow_threading::AllowThreads;
@@ -151,7 +151,7 @@ pub(crate) mod ext {
 /// 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.
-#[pymodule(name = "exo_pyo3_bindings")]
+#[pymodule(name = "exo_net")]
 fn main_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
     // install logger
     pyo3_log::init();
@@ -162,7 +162,12 @@ 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>()?;
+    // m.add_class::<PyKeypair>()?;
+    // networking_submodule(m)?;
+    m.add_class::<StateProxy>()?;
+    m.add_class::<PySession>()?;
+    m.add_class::<NetReceiver>()?;
+    m.add_class::<NetSender>()?;
     networking_submodule(m)?;
 
     // top-level constructs

rust/exo_net/src/networking.rs

@@ -0,0 +1,192 @@
+use std::pin::Pin;
+use std::sync::Arc;
+
+use crate::ext::{ByteArrayExt as _, FutureExt, PyErrExt as _};
+use crate::ext::{ResultExt as _, TokioMpscSenderExt as _};
+use crate::session::PySession;
+use futures_lite::{Stream, StreamExt as _};
+use networking::swarm::{FromSwarm, ToSwarm, create_swarm};
+use pyo3::exceptions::{PyRuntimeError, PyValueError};
+use pyo3::prelude::*;
+use pyo3::types::PyBytes;
+use pyo3::{Bound, Py, PyAny, PyErr, PyResult, Python, pymethods};
+use pyo3_stub_gen::derive::{
+    gen_methods_from_python, gen_stub_pyclass, gen_stub_pyclass_complex_enum, gen_stub_pymethods,
+};
+use tokio::sync::{Mutex, mpsc, oneshot};
+
+#[gen_stub_pyclass]
+#[pyclass(name = "NetworkingHandle")]
+struct PyNetworkingHandle {
+    // channels
+    pub to_swarm: mpsc::Sender<ToSwarm>,
+    pub swarm: Arc<Mutex<Pin<Box<dyn Stream<Item = FromSwarm> + Send>>>>,
+}
+
+#[gen_stub_pyclass_complex_enum]
+#[pyclass]
+enum PyFromSwarm {
+    Connection { connected: bool },
+    Message { topic: String, data: Py<PyBytes> },
+}
+impl From<FromSwarm> for PyFromSwarm {
+    fn from(value: FromSwarm) -> Self {
+        match value {
+            FromSwarm::Discovered {} => Self::Connection { connected: true },
+            FromSwarm::Expired {} => Self::Connection { connected: false },
+            FromSwarm::Message { topic, data } => Self::Message {
+                topic: topic,
+                data: data.pybytes(),
+            },
+        }
+    }
+}
+
+#[gen_stub_pymethods]
+#[pymethods]
+impl PyNetworkingHandle {
+    // NOTE: `async fn`s here that use `.await` will wrap the future in `.allow_threads_py()`
+    //       immediately beforehand to release the interpreter.
+    //       SEE: https://pyo3.rs/v0.26.0/async-await.html#detaching-from-the-interpreter-across-await
+
+    // ---- Lifecycle management methods ----
+
+    #[staticmethod]
+    fn new<'py>(
+        identity: Bound<'py, PyBytes>,
+        bootstrap_peers: Vec<String>,
+        listen_port: u16,
+    ) -> PyResult<(PyNetworkingHandle, PySession)> {
+        // create communication channels
+        let (to_swarm, from_client) = mpsc::channel(1024);
+
+        // get identity
+        let identity = u128::from_le_bytes(
+            identity
+                .extract::<'_, Vec<u8>>()?
+                .try_into()
+                .map_err(|_| PyValueError::new_err("invalid identity bytes"))?,
+        );
+
+        // create networking swarm (within tokio context!! or it crashes)
+        let swarm = pyo3_async_runtimes::tokio::get_runtime()
+            .block_on(create_swarm(
+                identity,
+                from_client,
+                bootstrap_peers,
+                listen_port,
+            ))
+            .pyerr()?;
+
+        let session = swarm.session.z.clone();
+
+        Ok((
+            PyNetworkingHandle {
+                swarm: Arc::new(Mutex::new(swarm.into_stream())),
+                to_swarm,
+            },
+            PySession { session },
+        ))
+    }
+
+    #[gen_stub(skip)]
+    fn recv<'py>(&'py self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
+        let swarm = Arc::clone(&self.swarm);
+        pyo3_async_runtimes::tokio::future_into_py(py, async move {
+            swarm
+                .try_lock()
+                .map_err(|_| PyRuntimeError::new_err("called recv twice concurrently"))?
+                .next()
+                .await
+                .ok_or(PyErr::receiver_channel_closed())
+                .map(PyFromSwarm::from)
+        })
+    }
+
+    // ---- Gossipsub management methods ----
+
+    /// Subscribe to a `GossipSub` topic.
+    ///
+    /// Returns `True` if the subscription worked. Returns `False` if we were already subscribed.
+    async fn gossipsub_subscribe(&self, topic: String) -> PyResult<bool> {
+        let (tx, rx) = oneshot::channel();
+
+        // send off request to subscribe
+        self.to_swarm
+            .send_py(ToSwarm::Subscribe {
+                topic,
+                result_sender: tx,
+            })
+            .allow_threads_py() // allow-threads-aware async call
+            .await?;
+
+        // wait for response & return any errors
+        rx.allow_threads_py() // allow-threads-aware async call
+            .await
+            .map_err(|_| PyErr::receiver_channel_closed())?
+            .pyerr()
+    }
+
+    /// Unsubscribes from a `GossipSub` topic.
+    ///
+    /// Returns `True` if we were subscribed to this topic. Returns `False` if we were not subscribed.
+    async fn gossipsub_unsubscribe(&self, topic: String) -> PyResult<bool> {
+        let (tx, rx) = oneshot::channel();
+
+        // send off request to unsubscribe
+        self.to_swarm
+            .send_py(ToSwarm::Unsubscribe {
+                topic,
+                result_sender: tx,
+            })
+            .allow_threads_py() // allow-threads-aware async call
+            .await?;
+
+        // wait for response & convert any errors
+        rx.allow_threads_py() // allow-threads-aware async call
+            .await
+            .map_err(|_| PyErr::receiver_channel_closed())
+    }
+
+    /// Publishes a message with multiple topics to the `GossipSub` network.
+    ///
+    /// If no peers are found that subscribe to this topic, throws `NoPeersSubscribedToTopicError` exception.
+    async fn gossipsub_publish(&self, topic: String, data: Py<PyBytes>) -> PyResult<()> {
+        let (tx, rx) = oneshot::channel();
+
+        // send off request to subscribe
+        let data = Python::attach(|py| Vec::from(data.as_bytes(py)));
+        self.to_swarm
+            .send_py(ToSwarm::Publish {
+                topic,
+                data,
+                result_sender: tx,
+            })
+            .allow_threads_py() // allow-threads-aware async call
+            .await?;
+
+        // wait for response & return any errors => ignore messageID for now!!!
+        let _ = rx
+            .allow_threads_py() // allow-threads-aware async call
+            .await
+            .map_err(|_| PyErr::receiver_channel_closed())?
+            .map_err(|e| PyRuntimeError::new_err(e.to_string()))?;
+        Ok(())
+    }
+}
+
+pyo3_stub_gen::inventory::submit! {
+    gen_methods_from_python! {
+        r#"
+            class PyNetworkingHandle:
+                async def recv() -> PyFromSwarm: ...
+        "#
+    }
+}
+
+pub fn networking_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
+    m.add_class::<PyNetworkingHandle>()?;
+    m.add_class::<PyFromSwarm>()?;
+
+    Ok(())
+}

rust/exo_net/src/point_to_point.rs

@@ -0,0 +1,108 @@
+use std::sync::Arc;
+
+use pyo3::exceptions::PyConnectionError;
+use pyo3::types::PyBytes;
+use pyo3::types::PyNone;
+use pyo3::{BoundObject, prelude::*};
+use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
+use zenoh::Result;
+use zenoh::{
+    handlers::FifoChannelHandler,
+    pubsub::{Publisher, Subscriber},
+    sample::Sample,
+};
+
+use crate::ext::ByteArrayExt;
+
+#[gen_stub_pyclass]
+#[pyclass]
+pub struct NetReceiver {
+    pub subscriber: Subscriber<FifoChannelHandler<Sample>>,
+}
+#[gen_stub_pymethods]
+#[pymethods]
+impl NetReceiver {
+    #[gen_stub(override_return_type(
+        type_repr="collections.abc.Awaitable[bytes | None]",
+        imports=("collections.abc")
+    ))]
+    pub fn recv<'py>(&'py self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
+        pyo3_async_runtimes::tokio::future_into_py(py, {
+            assert!(
+                self.subscriber.receiver_count() == 1,
+                "tried to receive twice on the same receiver"
+            );
+            let subscriber = self.subscriber.clone();
+            async move {
+                match subscriber.recv_async().await {
+                    Err(_) => {
+                        // stream closed;
+                        Ok(Python::attach(|py| PyNone::get(py).unbind()).into_any())
+                    }
+                    Ok(sample) => Ok(sample.payload().to_bytes().to_vec().pybytes().into_any()),
+                }
+            }
+        })
+    }
+}
+
+#[gen_stub_pyclass]
+#[pyclass]
+pub struct NetSender {
+    pub publisher: Arc<Publisher<'static>>,
+    pub first: bool,
+}
+#[gen_stub_pymethods]
+#[pymethods]
+impl NetSender {
+    #[gen_stub(override_return_type(
+        type_repr="collections.abc.Awaitable[bool]",
+        imports=("collections.abc")
+    ))]
+    pub fn send<'py>(
+        &'py mut self,
+        py: Python<'py>,
+        data: Bound<'py, PyBytes>,
+    ) -> PyResult<Bound<'py, PyAny>> {
+        let is_first = self.first;
+        self.first = false;
+        pyo3_async_runtimes::tokio::future_into_py(py, {
+            let publisher = Arc::clone(&self.publisher);
+            // clone the data so py can have it back
+            let bytes = data.as_bytes().to_vec();
+            async move {
+                if is_first {
+                    wait_for_listener(&*publisher)
+                        .await
+                        .map_err(|e| PyConnectionError::new_err(e.to_string()))?;
+                }
+                if !publisher
+                    .matching_status()
+                    .await
+                    .map_err(|e| PyConnectionError::new_err(e.to_string()))?
+                    .matching()
+                {
+                    return Ok(false);
+                }
+                publisher
+                    .put(&bytes)
+                    .await
+                    .map_err(|e| PyConnectionError::new_err(e.to_string()))?;
+                Ok(true)
+            }
+        })
+    }
+}
+
+async fn wait_for_listener<'a>(publisher: &Publisher<'a>) -> Result<()> {
+    let matcher = publisher.matching_listener().await?;
+    if publisher.matching_status().await?.matching() {
+        return Ok(());
+    }
+    while let Ok(status) = matcher.recv_async().await {
+        if status.matching() {
+            break;
+        }
+    }
+    Ok(())
+}

rust/exo_net/src/session.rs

@@ -0,0 +1,72 @@
+use std::sync::Arc;
+
+use pyo3::{exceptions::PyValueError, prelude::*};
+use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
+
+use zenoh::Session;
+use zenoh::Wait;
+use zenoh::qos::CongestionControl;
+
+use crate::{
+    point_to_point::{NetReceiver, NetSender},
+    state::StateProxy,
+};
+
+#[gen_stub_pyclass]
+#[pyclass]
+pub struct PySession {
+    pub session: Session,
+}
+#[gen_stub_pymethods]
+#[pymethods]
+impl PySession {
+    /* for now construct with NetworkingHandle
+    #[staticmethod]
+    #[gen_stub(override_return_type(
+        type_repr="collections.abc.Awaitable[PySession]",
+        imports=("collections.abc")
+    ))]
+    pub fn init<'py>(py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
+        pyo3_async_runtimes::tokio::future_into_py(py, async move {
+            Ok(Self {
+                session: networking::open(
+                    networking::cfg(rand::random(), 0).expect("default cfg is valid"),
+                )
+                .await
+                .map_err(|e| PyRuntimeError::new_err(e.to_string()))?,
+            })
+        })
+    }
+    */
+
+    pub fn net_receiver<'py>(&self, key: String) -> PyResult<NetReceiver> {
+        Ok(NetReceiver {
+            subscriber: self
+                .session
+                .declare_subscriber(key)
+                .wait()
+                // C5: key format error
+                .map_err(|e| PyValueError::new_err(e.to_string()))?,
+        })
+    }
+
+    pub fn net_sender<'py>(&self, key: String) -> PyResult<NetSender> {
+        Ok(NetSender {
+            publisher: Arc::new(
+                self.session
+                    .declare_publisher(key)
+                    .congestion_control(CongestionControl::Block)
+                    .wait()
+                    // C5: key format error, could be declaration error
+                    .map_err(|e| PyValueError::new_err(e.to_string()))?,
+            ),
+            first: true,
+        })
+    }
+
+    pub fn state_proxy(&self) -> StateProxy {
+        StateProxy {
+            session: self.session.clone(),
+        }
+    }
+}

rust/exo_net/src/state.rs

@@ -0,0 +1,78 @@
+use pyo3::{exceptions::PyValueError, prelude::*};
+use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
+use serde_json::{Map, Value};
+use zenoh::{Result, Session, sample::SampleFields};
+
+#[gen_stub_pyclass]
+#[pyclass]
+pub struct StateProxy {
+    pub session: Session,
+}
+#[gen_stub_pymethods]
+#[pymethods]
+impl StateProxy {
+    #[gen_stub(override_return_type(
+        type_repr="collections.abc.Awaitable[str]",
+        imports=("collections.abc")
+    ))]
+    pub fn snapshot<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
+        pyo3_async_runtimes::tokio::future_into_py(py, {
+            let session = self.session.clone();
+            async move {
+                Self::_snapshot(session)
+                    .await
+                    .map_err(|e| PyValueError::new_err(e.to_string()))
+                    .map(|v| v.to_string())
+            }
+        })
+    }
+}
+
+impl StateProxy {
+    async fn _snapshot(session: Session) -> Result<Value> {
+        let q = session.get("storage/mem1/**").await?;
+
+        let mut v = Value::Object(Map::default());
+
+        while let Ok(sample) = q.recv_async().await {
+            let mut cur_v = &mut v;
+            let Ok(sample) = sample.into_result() else {
+                continue;
+            };
+            // skip storage/mem1
+            let SampleFields {
+                payload, key_expr, ..
+            } = sample.into();
+            let mut iter = key_expr.split('/').skip(2).peekable();
+            loop {
+                let Some(p) = iter.next() else {
+                    break;
+                };
+                if iter.peek().is_none() {
+                    // terminal; write value into json
+                    let existing = cur_v
+                        .as_object_mut()
+                        .expect("path terminated unexpectedly - value stored at some/path and some/path/two")
+                        .insert(p.to_owned(), Value::String(payload.try_to_string()?.to_string()));
+
+                    if let Some(value) = existing {
+                        assert!(value.is_string())
+                        // could log, but string overwrites are fine
+                    }
+                } else {
+                    // non-terminal; ensure key exists in v, then replace cur with that object
+                    cur_v = cur_v
+                        .as_object_mut()
+                        .expect("path terminated unexpectedly - value stored at some/path and some/path/two")
+                        .entry(p)
+                        .or_insert(Value::Object(Map::default()));
+                    assert!(
+                        cur_v.is_object(),
+                        "path terminated unexpectedly - value stored at some/path and some/path/two"
+                    )
+                }
+            }
+        }
+        Ok(v)
+    }
+}

rust/exo_pyo3_bindings/Cargo.toml

@@ -1,66 +0,0 @@
-[package]
-name = "exo_pyo3_bindings"
-version = { workspace = true }
-edition = { workspace = true }
-publish = false
-
-[lib]
-doctest = false
-path = "src/lib.rs"
-name = "exo_pyo3_bindings"
-
-# "cdylib" needed to produce shared library for Python to import
-# "rlib" needed for stub-gen to run
-crate-type = ["cdylib", "rlib"]
-
-[[bin]]
-path = "src/bin/stub_gen.rs"
-name = "stub_gen"
-doc = false
-
-[lints]
-workspace = true
-
-[dependencies]
-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
-  "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
-
-  # integrations with other libraries
-  # "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" }
-pyo3-async-runtimes = { version = "0.27.0", features = [
-  "attributes",
-  "tokio-runtime",
-  "testing",
-] }
-pyo3-log = "0.13.2"
-
-# macro dependencies
-extend = { workspace = true }
-delegate = { workspace = true }
-
-# async runtime
-tokio = { workspace = true, features = ["full", "tracing"] }
-futures-lite = { workspace = true }
-
-# utility dependencies
-util = { workspace = true }
-
-# Tracing
-log = { workspace = true }
-env_logger = "0.11"
-
-# Networking
-libp2p = { workspace = true, features = ["full"] }
-pin-project = "1.1.10"

rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi

@@ -1,94 +0,0 @@
-# This file is automatically generated by pyo3_stub_gen
-# ruff: noqa: E501, F401
-
-import builtins
-import typing
-
-@typing.final
-class AllQueuesFullError(builtins.Exception):
-    def __new__(cls, *args: typing.Any) -> AllQueuesFullError: ...
-    def __repr__(self) -> builtins.str: ...
-    def __str__(self) -> builtins.str: ...
-
-@typing.final
-class Keypair:
-    r"""
-    Identity keypair of a node.
-    """
-    @staticmethod
-    def generate() -> Keypair:
-        r"""
-        Generate a new Ed25519 keypair.
-        """
-    @staticmethod
-    def from_bytes(bytes: bytes) -> Keypair:
-        r"""
-        Construct an Ed25519 keypair from secret key bytes
-        """
-    def to_bytes(self) -> bytes:
-        r"""
-        Get the secret key bytes underlying the keypair
-        """
-    def to_node_id(self) -> builtins.str:
-        r"""
-        Convert the `Keypair` into the corresponding `PeerId` string, which we use as our `NodeId`.
-        """
-
-@typing.final
-class MessageTooLargeError(builtins.Exception):
-    def __new__(cls, *args: typing.Any) -> MessageTooLargeError: ...
-    def __repr__(self) -> builtins.str: ...
-    def __str__(self) -> builtins.str: ...
-
-@typing.final
-class NetworkingHandle:
-    def __new__(cls, identity: Keypair, bootstrap_peers: typing.Sequence[builtins.str], listen_port: builtins.int) -> NetworkingHandle: ...
-    async def gossipsub_subscribe(self, topic: builtins.str) -> builtins.bool:
-        r"""
-        Subscribe to a `GossipSub` topic.
-        
-        Returns `True` if the subscription worked. Returns `False` if we were already subscribed.
-        """
-    async def gossipsub_unsubscribe(self, topic: builtins.str) -> builtins.bool:
-        r"""
-        Unsubscribes from a `GossipSub` topic.
-        
-        Returns `True` if we were subscribed to this topic. Returns `False` if we were not subscribed.
-        """
-    async def gossipsub_publish(self, topic: builtins.str, data: bytes) -> None:
-        r"""
-        Publishes a message with multiple topics to the `GossipSub` network.
-        
-        If no peers are found that subscribe to this topic, throws `NoPeersSubscribedToTopicError` exception.
-        """
-    async def recv(self) -> PyFromSwarm: ...
-
-@typing.final
-class NoPeersSubscribedToTopicError(builtins.Exception):
-    def __new__(cls, *args: typing.Any) -> NoPeersSubscribedToTopicError: ...
-    def __repr__(self) -> builtins.str: ...
-    def __str__(self) -> builtins.str: ...
-
-class PyFromSwarm:
-    @typing.final
-    class Connection(PyFromSwarm):
-        __match_args__ = ("peer_id", "connected",)
-        @property
-        def peer_id(self) -> builtins.str: ...
-        @property
-        def connected(self) -> builtins.bool: ...
-        def __new__(cls, peer_id: builtins.str, connected: builtins.bool) -> PyFromSwarm.Connection: ...
-    
-    @typing.final
-    class Message(PyFromSwarm):
-        __match_args__ = ("origin", "topic", "data",)
-        @property
-        def origin(self) -> builtins.str: ...
-        @property
-        def topic(self) -> builtins.str: ...
-        @property
-        def data(self) -> bytes: ...
-        def __new__(cls, origin: builtins.str, topic: builtins.str, data: bytes) -> PyFromSwarm.Message: ...
-    
-    ...
-

rust/exo_pyo3_bindings/src/ident.rs

@@ -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()
-    }
-}

rust/exo_pyo3_bindings/src/networking.rs

@@ -1,318 +0,0 @@
-use std::pin::Pin;
-use std::sync::Arc;
-
-use crate::r#const::MPSC_CHANNEL_SIZE;
-use crate::ext::{ByteArrayExt as _, FutureExt, PyErrExt as _};
-use crate::ext::{ResultExt as _, TokioMpscSenderExt as _};
-use crate::ident::PyKeypair;
-use crate::networking::exception::{
-    PyAllQueuesFullError, PyMessageTooLargeError, PyNoPeersSubscribedToTopicError,
-};
-use crate::pyclass;
-use futures_lite::{Stream, StreamExt as _};
-use libp2p::gossipsub::PublishError;
-use networking::swarm::{FromSwarm, ToSwarm, create_swarm};
-use pyo3::exceptions::PyRuntimeError;
-use pyo3::prelude::{PyModule, PyModuleMethods as _};
-use pyo3::types::PyBytes;
-use pyo3::{Bound, Py, PyAny, PyErr, PyResult, Python, pymethods};
-use pyo3_stub_gen::derive::{
-    gen_methods_from_python, gen_stub_pyclass, gen_stub_pyclass_complex_enum, gen_stub_pymethods,
-};
-use tokio::sync::{Mutex, mpsc, oneshot};
-
-mod exception {
-    use pyo3::types::PyTuple;
-    use pyo3::{exceptions::PyException, prelude::*};
-    use pyo3_stub_gen::derive::*;
-
-    #[gen_stub_pyclass]
-    #[pyclass(frozen, extends=PyException, name="NoPeersSubscribedToTopicError")]
-    pub struct PyNoPeersSubscribedToTopicError {}
-
-    impl PyNoPeersSubscribedToTopicError {
-        const MSG: &'static str = "\
-        No peers are currently subscribed to receive messages on this topic. \
-        Wait for peers to subscribe or check your network connectivity.";
-
-        ///   Creates a new  [ `PyErr` ]  of this type.
-        ///
-        ///   [`PyErr`] :  https://docs.rs/pyo3/latest/pyo3/struct.PyErr.html   "PyErr in pyo3"
-        pub(crate) fn new_err() -> PyErr {
-            PyErr::new::<Self, _>(()) // TODO: check if this needs to be replaced???
-        }
-    }
-
-    #[gen_stub_pymethods]
-    #[pymethods]
-    impl PyNoPeersSubscribedToTopicError {
-        #[new]
-        #[pyo3(signature = (*args))]
-        #[allow(unused_variables)]
-        pub(crate) fn new(args: &Bound<'_, PyTuple>) -> Self {
-            Self {}
-        }
-
-        fn __repr__(&self) -> String {
-            format!("PeerId(\"{}\")", Self::MSG)
-        }
-
-        fn __str__(&self) -> String {
-            Self::MSG.to_string()
-        }
-    }
-
-    #[gen_stub_pyclass]
-    #[pyclass(frozen, extends=PyException, name="AllQueuesFullError")]
-    pub struct PyAllQueuesFullError {}
-
-    impl PyAllQueuesFullError {
-        const MSG: &'static str =
-            "All libp2p peers are unresponsive, resend the message or reconnect.";
-
-        ///   Creates a new  [ `PyErr` ]  of this type.
-        ///
-        ///   [`PyErr`] :  https://docs.rs/pyo3/latest/pyo3/struct.PyErr.html   "PyErr in pyo3"
-        pub(crate) fn new_err() -> PyErr {
-            PyErr::new::<Self, _>(()) // TODO: check if this needs to be replaced???
-        }
-    }
-
-    #[gen_stub_pymethods]
-    #[pymethods]
-    impl PyAllQueuesFullError {
-        #[new]
-        #[pyo3(signature = (*args))]
-        #[allow(unused_variables)]
-        pub(crate) fn new(args: &Bound<'_, PyTuple>) -> Self {
-            Self {}
-        }
-
-        fn __repr__(&self) -> String {
-            format!("PeerId(\"{}\")", Self::MSG)
-        }
-
-        fn __str__(&self) -> String {
-            Self::MSG.to_string()
-        }
-    }
-
-    #[gen_stub_pyclass]
-    #[pyclass(frozen, extends=PyException, name="MessageTooLargeError")]
-    pub struct PyMessageTooLargeError {}
-
-    impl PyMessageTooLargeError {
-        const MSG: &'static str = "Gossipsub message exceeds max_transmit_size. Reduce prompt length or increase the limit.";
-
-        pub(crate) fn new_err() -> PyErr {
-            PyErr::new::<Self, _>(())
-        }
-    }
-
-    #[gen_stub_pymethods]
-    #[pymethods]
-    impl PyMessageTooLargeError {
-        #[new]
-        #[pyo3(signature = (*args))]
-        #[allow(unused_variables)]
-        pub(crate) fn new(args: &Bound<'_, PyTuple>) -> Self {
-            Self {}
-        }
-
-        fn __repr__(&self) -> String {
-            format!("MessageTooLargeError(\"{}\")", Self::MSG)
-        }
-
-        fn __str__(&self) -> String {
-            Self::MSG.to_string()
-        }
-    }
-}
-
-#[gen_stub_pyclass]
-#[pyclass(name = "NetworkingHandle")]
-struct PyNetworkingHandle {
-    // channels
-    pub to_swarm: mpsc::Sender<ToSwarm>,
-    pub swarm: Arc<Mutex<Pin<Box<dyn Stream<Item = FromSwarm> + Send>>>>,
-}
-
-#[gen_stub_pyclass_complex_enum]
-#[pyclass]
-enum PyFromSwarm {
-    Connection {
-        peer_id: String,
-        connected: bool,
-    },
-    Message {
-        origin: String,
-        topic: String,
-        data: Py<PyBytes>,
-    },
-}
-impl From<FromSwarm> for PyFromSwarm {
-    fn from(value: FromSwarm) -> Self {
-        match value {
-            FromSwarm::Discovered { peer_id } => Self::Connection {
-                peer_id: peer_id.to_base58(),
-                connected: true,
-            },
-            FromSwarm::Expired { peer_id } => Self::Connection {
-                peer_id: peer_id.to_base58(),
-                connected: false,
-            },
-            FromSwarm::Message { from, topic, data } => Self::Message {
-                origin: from.to_base58(),
-                topic: topic,
-                data: data.pybytes(),
-            },
-        }
-    }
-}
-
-#[gen_stub_pymethods]
-#[pymethods]
-impl PyNetworkingHandle {
-    // NOTE: `async fn`s here that use `.await` will wrap the future in `.allow_threads_py()`
-    //       immediately beforehand to release the interpreter.
-    //       SEE: https://pyo3.rs/v0.26.0/async-await.html#detaching-from-the-interpreter-across-await
-
-    // ---- Lifecycle management methods ----
-
-    #[new]
-    #[pyo3(signature = (identity, bootstrap_peers, listen_port))]
-    fn py_new(
-        identity: Bound<'_, PyKeypair>,
-        bootstrap_peers: Vec<String>,
-        listen_port: u16,
-    ) -> PyResult<Self> {
-        // create communication channels
-        let (to_swarm, from_client) = mpsc::channel(MPSC_CHANNEL_SIZE);
-
-        // get identity
-        let identity = identity.borrow().0.clone();
-
-        // create networking swarm (within tokio context!! or it crashes)
-        let _guard = pyo3_async_runtimes::tokio::get_runtime().enter();
-        let swarm = create_swarm(identity, from_client, bootstrap_peers, listen_port)
-            .pyerr()?
-            .into_stream();
-
-        Ok(Self {
-            swarm: Arc::new(Mutex::new(swarm)),
-            to_swarm,
-        })
-    }
-
-    #[gen_stub(skip)]
-    fn recv<'py>(&'py self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
-        let swarm = Arc::clone(&self.swarm);
-        pyo3_async_runtimes::tokio::future_into_py(py, async move {
-            swarm
-                .try_lock()
-                .map_err(|_| PyRuntimeError::new_err("called recv twice concurrently"))?
-                .next()
-                .await
-                .ok_or(PyErr::receiver_channel_closed())
-                .map(PyFromSwarm::from)
-        })
-    }
-
-    // ---- Gossipsub management methods ----
-
-    /// Subscribe to a `GossipSub` topic.
-    ///
-    /// Returns `True` if the subscription worked. Returns `False` if we were already subscribed.
-    async fn gossipsub_subscribe(&self, topic: String) -> PyResult<bool> {
-        let (tx, rx) = oneshot::channel();
-
-        // send off request to subscribe
-        self.to_swarm
-            .send_py(ToSwarm::Subscribe {
-                topic,
-                result_sender: tx,
-            })
-            .allow_threads_py() // allow-threads-aware async call
-            .await?;
-
-        // wait for response & return any errors
-        rx.allow_threads_py() // allow-threads-aware async call
-            .await
-            .map_err(|_| PyErr::receiver_channel_closed())?
-            .pyerr()
-    }
-
-    /// Unsubscribes from a `GossipSub` topic.
-    ///
-    /// Returns `True` if we were subscribed to this topic. Returns `False` if we were not subscribed.
-    async fn gossipsub_unsubscribe(&self, topic: String) -> PyResult<bool> {
-        let (tx, rx) = oneshot::channel();
-
-        // send off request to unsubscribe
-        self.to_swarm
-            .send_py(ToSwarm::Unsubscribe {
-                topic,
-                result_sender: tx,
-            })
-            .allow_threads_py() // allow-threads-aware async call
-            .await?;
-
-        // wait for response & convert any errors
-        rx.allow_threads_py() // allow-threads-aware async call
-            .await
-            .map_err(|_| PyErr::receiver_channel_closed())
-    }
-
-    /// Publishes a message with multiple topics to the `GossipSub` network.
-    ///
-    /// If no peers are found that subscribe to this topic, throws `NoPeersSubscribedToTopicError` exception.
-    async fn gossipsub_publish(&self, topic: String, data: Py<PyBytes>) -> PyResult<()> {
-        let (tx, rx) = oneshot::channel();
-
-        // send off request to subscribe
-        let data = Python::attach(|py| Vec::from(data.as_bytes(py)));
-        self.to_swarm
-            .send_py(ToSwarm::Publish {
-                topic,
-                data,
-                result_sender: tx,
-            })
-            .allow_threads_py() // allow-threads-aware async call
-            .await?;
-
-        // wait for response & return any errors => ignore messageID for now!!!
-        let _ = rx
-            .allow_threads_py() // allow-threads-aware async call
-            .await
-            .map_err(|_| PyErr::receiver_channel_closed())?
-            .map_err(|e| match e {
-                PublishError::AllQueuesFull(_) => PyAllQueuesFullError::new_err(),
-                PublishError::MessageTooLarge => PyMessageTooLargeError::new_err(),
-                PublishError::NoPeersSubscribedToTopic => {
-                    PyNoPeersSubscribedToTopicError::new_err()
-                }
-                e => PyRuntimeError::new_err(e.to_string()),
-            })?;
-        Ok(())
-    }
-}
-
-pyo3_stub_gen::inventory::submit! {
-    gen_methods_from_python! {
-        r#"
-            class PyNetworkingHandle:
-                async def recv() -> PyFromSwarm: ...
-        "#
-    }
-}
-
-pub fn networking_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
-    m.add_class::<exception::PyNoPeersSubscribedToTopicError>()?;
-    m.add_class::<exception::PyAllQueuesFullError>()?;
-    m.add_class::<exception::PyMessageTooLargeError>()?;
-
-    m.add_class::<PyNetworkingHandle>()?;
-    m.add_class::<PyFromSwarm>()?;
-
-    Ok(())
-}

rust/exo_pyo3_bindings/tests/dummy.rs

@@ -1,54 +0,0 @@
-#[cfg(test)]
-mod tests {
-    use core::mem::drop;
-    use core::option::Option::Some;
-    use core::time::Duration;
-    use tokio;
-    use tokio::sync::mpsc;
-
-    #[tokio::test]
-    async fn test_drop_channel() {
-        struct Ping;
-
-        let (tx, mut rx) = mpsc::channel::<Ping>(10);
-
-        let _ = tokio::spawn(async move {
-            println!("TASK: entered");
-
-            loop {
-                tokio::select! {
-                    result = rx.recv() => {
-                        match result {
-                            Some(_) => {
-                                println!("TASK: pinged");
-                            }
-                            None => {
-                                println!("TASK: closing channel");
-                                break;
-                            }
-                        }
-                    }
-                    _ = tokio::time::sleep(Duration::from_secs_f32(0.1)) => {
-                        println!("TASK: heartbeat");
-                    }
-                }
-            }
-
-            println!("TASK: exited");
-        });
-
-        let tx2 = tx.clone();
-
-        tokio::time::sleep(Duration::from_secs_f32(0.11)).await;
-
-        tx.send(Ping).await.expect("Should not fail");
-        drop(tx);
-
-        tokio::time::sleep(Duration::from_secs_f32(0.11)).await;
-
-        tx2.send(Ping).await.expect("Should not fail");
-        drop(tx2);
-
-        tokio::time::sleep(Duration::from_secs_f32(0.11)).await;
-    }
-}

rust/exo_pyo3_bindings/tests/test_python.py

@@ -1,36 +0,0 @@
-import asyncio
-
-import pytest
-from exo_pyo3_bindings import (
-    Keypair,
-    NetworkingHandle,
-    NoPeersSubscribedToTopicError,
-    PyFromSwarm,
-)
-
-
-@pytest.mark.asyncio
-async def test_sleep_on_multiple_items() -> None:
-    print("PYTHON: starting handle")
-    h = NetworkingHandle(Keypair.generate(), [], 0)
-
-    rt = asyncio.create_task(_await_recv(h))
-
-    # sleep for 4 ticks
-    for i in range(4):
-        await asyncio.sleep(1)
-
-        try:
-            await h.gossipsub_publish("topic", b"somehting or other")
-        except NoPeersSubscribedToTopicError as e:
-            print("caught it", e)
-
-
-async def _await_recv(h: NetworkingHandle):
-    while True:
-        event = await h.recv()
-        match event:
-            case PyFromSwarm.Connection() as c:
-                print(f"PYTHON: connection update: {c}")
-            case PyFromSwarm.Message() as m:
-                print(f"PYTHON: message: {m}")

rust/networking/Cargo.toml

@@ -1,42 +1,20 @@
 [package]
 name = "networking"
-version = { workspace = true }
-edition = { workspace = true }
-publish = false
+version.workspace = true
+edition.workspace = true
 
-[lib]
-doctest = false
-name = "networking"
-path = "src/lib.rs"
+[dependencies]
+async-stream.workspace = true
+futures-lite.workspace = true
+netwatcher = { workspace = true, features = ["tokio"] }
+parking_lot.workspace = true
+tokio = { workspace = true, features = ["full"] }
+zenoh = { workspace = true, features = ["internal", "plugins", "unstable"] }
+zenoh-plugin-storage-manager.workspace = true
+zenoh-plugin-trait.workspace = true
+rand.workspace = true
+log.workspace = true
+bytemuck = { workspace = true, features = ["derive"] }
 
 [lints]
 workspace = true
-
-[dependencies]
-# datastructures
-either = { workspace = true }
-
-# macro dependencies
-extend = { workspace = true }
-delegate = { workspace = true }
-
-# async
-async-stream = { workspace = true }
-futures-lite = { workspace = true }
-futures-timer = { workspace = true }
-tokio = { workspace = true, features = ["full"] }
-
-# utility dependencies
-util = { workspace = true }
-tracing-subscriber = { version = "0.3.19", features = [
-  "default",
-  "env-filter",
-] }
-keccak-const = { workspace = true }
-
-# tracing/logging
-log = { workspace = true }
-
-# networking
-libp2p = { workspace = true, features = ["full"] }
-pin-project = "1.1.10"

rust/networking/examples/chatroom.rs

@@ -1,86 +0,0 @@
-use futures_lite::StreamExt;
-use libp2p::identity;
-use networking::swarm;
-use networking::swarm::{FromSwarm, ToSwarm};
-use tokio::sync::{mpsc, oneshot};
-use tokio::{io, io::AsyncBufReadExt as _};
-use tracing_subscriber::EnvFilter;
-use tracing_subscriber::filter::LevelFilter;
-
-#[tokio::main]
-async fn main() {
-    let _ = tracing_subscriber::fmt()
-        .with_env_filter(EnvFilter::from_default_env().add_directive(LevelFilter::INFO.into()))
-        .try_init();
-
-    let (to_swarm, from_client) = mpsc::channel(20);
-
-    // Configure swarm
-    let mut swarm = swarm::create_swarm(
-        identity::Keypair::generate_ed25519(),
-        from_client,
-        vec![],
-        0,
-    )
-    .expect("Swarm creation failed")
-    .into_stream();
-
-    // Create a Gossipsub topic & subscribe
-    let (tx, rx) = oneshot::channel();
-    _ = to_swarm
-        .send(ToSwarm::Subscribe {
-            topic: "test-net".to_string(),
-            result_sender: tx,
-        })
-        .await
-        .expect("should send");
-
-    // Read full lines from stdin
-    let mut stdin = io::BufReader::new(io::stdin()).lines();
-    println!("Enter messages via STDIN and they will be sent to connected peers using Gossipsub");
-
-    tokio::task::spawn(async move {
-        rx.await
-            .expect("tx not dropped")
-            .expect("subscribe shouldn't fail");
-        loop {
-            if let Ok(Some(line)) = stdin.next_line().await {
-                let (tx, rx) = oneshot::channel();
-                if let Err(e) = to_swarm
-                    .send(swarm::ToSwarm::Publish {
-                        topic: "test-net".to_string(),
-                        data: line.as_bytes().to_vec(),
-                        result_sender: tx,
-                    })
-                    .await
-                {
-                    println!("Send error: {e:?}");
-                    return;
-                };
-                match rx.await {
-                    Ok(Err(e)) => println!("Publish error: {e:?}"),
-                    Err(e) => println!("Publish error: {e:?}"),
-                    Ok(_) => {}
-                }
-            }
-        }
-    });
-
-    // Kick it off
-    loop {
-        // on gossipsub outgoing
-        match swarm.next().await {
-            // on gossipsub incoming
-            Some(FromSwarm::Discovered { peer_id }) => {
-                println!("\n\nconnected to {peer_id}\n\n")
-            }
-            Some(FromSwarm::Expired { peer_id }) => {
-                println!("\n\ndisconnected from {peer_id}\n\n")
-            }
-            Some(FromSwarm::Message { from, topic, data }) => {
-                println!("{topic}/{from}:\n{}", String::from_utf8_lossy(&data))
-            }
-            None => {}
-        }
-    }
-}

rust/networking/examples/put_string.rs

@@ -0,0 +1,23 @@
+use log::info;
+use networking;
+use zenoh::Result;
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    zenoh::init_log_from_env_or("info");
+    info!("Opening session...");
+    let cfg = networking::cfg(rand::random(), 0)?;
+    let session = networking::open(cfg, 52414).await?;
+    let _tok = session
+        .z
+        .liveliness()
+        .declare_token(format!("nodes/{}/live", session.z.zid()))
+        .await?;
+    let key_expr = "storage/mem1/name";
+    let payload = "me";
+
+    info!("Putting Data ('{key_expr}': '{payload}')...");
+    session.z.put(key_expr, payload).await?;
+    tokio::signal::ctrl_c().await?;
+    Ok(())
+}

rust/networking/examples/serve_storage.rs

@@ -0,0 +1,25 @@
+use log::info;
+use networking;
+use zenoh::Result;
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    zenoh::init_log_from_env_or("info");
+    info!("Opening session...");
+    let cfg = networking::cfg(rand::random(), 52414)?;
+    let session = networking::open(cfg, 52414).await?;
+    let _tok = session
+        .z
+        .liveliness()
+        .declare_token(format!("nodes/{}/live", session.z.zid()))
+        .await?;
+    let _sub = session
+        .z
+        .liveliness()
+        .declare_subscriber("nodes/*/live")
+        .history(true)
+        .callback(|tok| println!("{tok:?}"))
+        .await?;
+    tokio::signal::ctrl_c().await?;
+    Ok(())
+}

rust/networking/src/RESEARCH_NOTES.txt

@@ -1,44 +0,0 @@
-https://github.com/ml-explore/mlx/commit/3fe98bacc7640d857acf3539f1d21b47a32e5609
-^raw sockets distributed -> `<net/ndrv.h>` -> https://newosxbook.com/code/xnu-3247.1.106/bsd/net/ndrv.h.auto.html
---> header file for a networking component found in the macOS kernel (XNU) that defines structures for network device driver registration, specifically the ndrv_demux_desc and ndrv_protocol_desc structures used for demultiplexing protocol data at the network interface level. It specifies how to describe protocol data, such as an Ethernet type or a SNAP header, and how to associate these descriptions with a specific protocol family to receive matching packets.
---> Used to bind an NDRV socket so that packets that match given protocol demux descriptions can be received.
---> An NDRV socket is a special kind of socket in the Darwin/macOS operating system's XNU kernel, used for low-level network packet manipulation and binding to specific protocols for packet processing. It allows user-space applications or drivers to directly write Layer 2 (L2) network packets or interact with the network stack at a lower level, often by binding to protocol descriptors like the ndrv_protocol_desc. This type of socket is used for functions such as capturing and injecting packets, especially in network infrastructure software like routers or for kernel-level network monitoring and security tools.
---> also called PF_NDRV sockets --> https://newosxbook.com/bonus/vol1ch16.html
-----> they are conceptually similar to https://scapy.disruptivelabs.in/networking/socket-interface PF_RAW or PF_PACKET
-
-https://stackoverflow.com/questions/17169298/af-packet-on-osx
-^AF_PACKET duplicates the packets as soon as it receives them from the physical layer (for incoming packets) or just before sending them out to the physical layer (for outgoing packets). -> this is on Linux only
-^it doesn't exist on OS X so you can use /dev/bpfX (Berkeley Packet Filter) for sniffing
-
-https://www.unix.com/man_page/mojave/4/ip/
-^OS X manpages for IP
-
-https://developer.apple.com/documentation/kernel/implementing_drivers_system_extensions_and_kexts
-^driver kit, system extensions & kexts for macOS
-
-----
-
-To set up a Linux system to use a Thunderbolt connection as a network device, connect the two computers with a Thunderbolt cable, load the thunderbolt-net kernel module (usually automatic but modprobe is an option for manual loading), and then the operating system will create virtual Ethernet interfaces (e.g., thunderbolt0) for networking. You can then use standard tools like ifconfig or your desktop environment's network manager to configure these new interfaces for a link-local network.
---> https://gist.github.com/geosp/80fbd39e617b7d1d9421683df4ea224a
-----> here is a guide on how to set up thunderbolt-ethernet on linux
-----> I may be able to steal the thunderbolt-net code ideas to implement a kernel module for MacOS
-
-https://chatgpt.com/s/t_68af8e41a8548191993281a014f846a7
-^GPT discussion about making socket interface
-
-https://chatgpt.com/s/t_68afb798a85c8191973c02a0fa7a48a3 --> link-local address,,??
-https://chatgpt.com/s/t_68afb02987e08191b2b0044d3667ece2
-^GPT discussion about accessing TB on MacOS low level interactions
-
---------------------------------
-
-https://www.intel.com/content/www/us/en/support/articles/000098893/software.html
-^Thunderbolt Share & Thunderbolt Networking Mode => intel's equivalent of thunderbolt bridge
-
-
----------------------------------
-
-https://www.zerotier.com/blog/how-zerotier-eliminated-kernel-extensions-on-macos/
--->fake ethernet devices on MacOS -> omg??? we can detect thunderbolt bridge, then bind to it, then re-expose it as fake ethernet??
--->ps: https://chatgpt.com/s/t_68afb2b25fb881919526763fb5d7359c, AF/PF_NDRV are one and the same!!!
--->https://github.com/zerotier/ZeroTierOne/blob/dev/osdep/MacEthernetTapAgent.c

rust/networking/src/discovery.rs

@@ -1,390 +1,311 @@
-use crate::ext::MultiaddrExt;
-use delegate::delegate;
-use either::Either;
-use futures_lite::FutureExt;
-use futures_timer::Delay;
-use libp2p::core::transport::PortUse;
-use libp2p::core::{ConnectedPoint, Endpoint};
-use libp2p::swarm::behaviour::ConnectionEstablished;
-use libp2p::swarm::dial_opts::DialOpts;
-use libp2p::swarm::{
-    CloseConnection, ConnectionClosed, ConnectionDenied, ConnectionHandler,
-    ConnectionHandlerSelect, ConnectionId, FromSwarm, NetworkBehaviour, THandler, THandlerInEvent,
-    THandlerOutEvent, ToSwarm, dummy,
+use std::{
+    io::{self, ErrorKind},
+    net::{Ipv6Addr, SocketAddr, SocketAddrV6},
+    sync::Arc,
+    time::Duration,
 };
-use libp2p::{Multiaddr, PeerId, identity, mdns};
-use std::collections::{BTreeSet, HashMap};
-use std::convert::Infallible;
-use std::io;
-use std::net::IpAddr;
-use std::task::{Context, Poll};
-use std::time::Duration;
-use util::wakerdeque::WakerDeque;
 
-const RETRY_CONNECT_INTERVAL: Duration = Duration::from_secs(5);
+use bytemuck::{Pod, Zeroable};
+use log::{debug, trace, warn};
+use netwatcher::WatchHandle;
+use parking_lot::Mutex;
+use tokio::{
+    net::UdpSocket,
+    time::{Interval, interval},
+};
+use zenoh::config::ZenohId;
 
-mod managed {
-    use libp2p::swarm::NetworkBehaviour;
-    use libp2p::{identity, mdns, ping};
-    use std::io;
-    use std::time::Duration;
+const GROUP: Ipv6Addr = Ipv6Addr::new(0xff12, 0, 0, 0, 0, 0, 0xe0a1, 0xde89);
 
-    const MDNS_RECORD_TTL: Duration = Duration::from_secs(2_500);
-    const MDNS_QUERY_INTERVAL: Duration = Duration::from_secs(1_500);
-    const PING_TIMEOUT: Duration = Duration::from_millis(2_500);
-    const PING_INTERVAL: Duration = Duration::from_millis(2_500);
-
-    #[derive(NetworkBehaviour)]
-    pub struct Behaviour {
-        mdns: mdns::tokio::Behaviour,
-        ping: ping::Behaviour,
-    }
-
-    impl Behaviour {
-        pub fn new(keypair: &identity::Keypair) -> io::Result<Self> {
-            Ok(Self {
-                mdns: mdns_behaviour(keypair)?,
-                ping: ping_behaviour(),
-            })
-        }
-    }
-
-    fn mdns_behaviour(keypair: &identity::Keypair) -> io::Result<mdns::tokio::Behaviour> {
-        use mdns::{Config, tokio};
-
-        // mDNS config => enable IPv6
-        let mdns_config = Config {
-            ttl: MDNS_RECORD_TTL,
-            query_interval: MDNS_QUERY_INTERVAL,
-
-            // enable_ipv6: true, // TODO: for some reason, TCP+mDNS don't work well with ipv6?? figure out how to make work
-            ..Default::default()
-        };
-
-        let mdns_behaviour = tokio::Behaviour::new(mdns_config, keypair.public().to_peer_id());
-        Ok(mdns_behaviour?)
-    }
-
-    fn ping_behaviour() -> ping::Behaviour {
-        ping::Behaviour::new(
-            ping::Config::new()
-                .with_timeout(PING_TIMEOUT)
-                .with_interval(PING_INTERVAL),
-        )
-    }
+pub struct Discovery {
+    sock: Arc<UdpSocket>,
+    ifaces: Arc<Mutex<Vec<SocketAddr>>>,
+    last_nonce: Mutex<[u8; 8]>,
+    /// the port of the service we are doing discovery for - transmitted to peers
+    listen_port: u16,
+    zid: ZenohId,
+    tick: Interval,
+    _sync: Mutex<WatchHandle>,
 }
 
-/// Events for when a listening connection is truly established and truly closed.
-#[derive(Debug, Clone)]
-pub enum Event {
-    ConnectionEstablished {
-        peer_id: PeerId,
-        connection_id: ConnectionId,
-        remote_ip: IpAddr,
-        remote_tcp_port: u16,
-    },
-    ConnectionClosed {
-        peer_id: PeerId,
-        connection_id: ConnectionId,
-        remote_ip: IpAddr,
-        remote_tcp_port: u16,
-    },
-}
-
-/// Discovery behavior that wraps mDNS to produce truly discovered durable peer-connections.
-///
-/// The behaviour operates as such:
-///  1) All true (listening) connections/disconnections are tracked, emitting corresponding events
-///     to the swarm.
-///  1) mDNS discovered/expired peers are tracked; discovered but not connected peers are dialed
-///     immediately, and expired but connected peers are disconnected from immediately.
-///  2) Every fixed interval: discovered but not connected peers are dialed, and expired but
-///     connected peers are disconnected from.
-pub struct Behaviour {
-    // state-tracking for managed behaviors & mDNS-discovered peers
-    managed: managed::Behaviour,
-    mdns_discovered: HashMap<PeerId, BTreeSet<Multiaddr>>,
-    bootstrap_peers: Vec<Multiaddr>,
-
-    retry_delay: Delay, // retry interval
-
-    // pending events to emmit => waker-backed Deque to control polling
-    pending_events: WakerDeque<ToSwarm<Event, Infallible>>,
-}
-
-impl Behaviour {
-    pub fn new(keypair: &identity::Keypair, bootstrap_peers: Vec<Multiaddr>) -> io::Result<Self> {
-        Ok(Self {
-            managed: managed::Behaviour::new(keypair)?,
-            mdns_discovered: HashMap::new(),
-            bootstrap_peers,
-            retry_delay: Delay::new(RETRY_CONNECT_INTERVAL),
-            pending_events: WakerDeque::new(),
-        })
-    }
-
-    fn dial(&mut self, peer_id: PeerId, addr: Multiaddr) {
-        self.pending_events.push_back(ToSwarm::Dial {
-            opts: DialOpts::peer_id(peer_id).addresses(vec![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 {
-            peer_id,
-            connection: CloseConnection::One(connection),
-        })
-    }
-
-    fn handle_mdns_discovered(&mut self, peers: Vec<(PeerId, Multiaddr)>) {
-        for (p, ma) in peers {
-            self.dial(p, ma.clone()); // always connect
-
-            // get peer's multi-addresses or insert if missing
-            let Some(mas) = self.mdns_discovered.get_mut(&p) else {
-                self.mdns_discovered.insert(p, BTreeSet::from([ma]));
-                continue;
-            };
-
-            // multiaddress should never already be present - else something has gone wrong
-            let is_new_addr = mas.insert(ma);
-            assert!(is_new_addr, "cannot discover a discovered peer");
-        }
-    }
-
-    fn handle_mdns_expired(&mut self, peers: Vec<(PeerId, Multiaddr)>) {
-        for (p, ma) in peers {
-            // at this point, we *must* have the peer
-            let mas = self
-                .mdns_discovered
-                .get_mut(&p)
-                .expect("nonexistent peer cannot expire");
-
-            // at this point, we *must* have the multiaddress
-            let was_present = mas.remove(&ma);
-            assert!(was_present, "nonexistent multiaddress cannot expire");
-
-            // if empty, remove the peer-id entirely
-            if mas.is_empty() {
-                self.mdns_discovered.remove(&p);
-            }
-        }
-    }
-
-    fn on_connection_established(
-        &mut self,
-        peer_id: PeerId,
-        connection_id: ConnectionId,
-        remote_ip: IpAddr,
-        remote_tcp_port: u16,
-    ) {
-        // send out connected event
-        self.pending_events
-            .push_back(ToSwarm::GenerateEvent(Event::ConnectionEstablished {
-                peer_id,
-                connection_id,
-                remote_ip,
-                remote_tcp_port,
-            }));
-    }
-
-    fn on_connection_closed(
-        &mut self,
-        peer_id: PeerId,
-        connection_id: ConnectionId,
-        remote_ip: IpAddr,
-        remote_tcp_port: u16,
-    ) {
-        // send out disconnected event
-        self.pending_events
-            .push_back(ToSwarm::GenerateEvent(Event::ConnectionClosed {
-                peer_id,
-                connection_id,
-                remote_ip,
-                remote_tcp_port,
-            }));
-    }
-}
-
-impl NetworkBehaviour for Behaviour {
-    type ConnectionHandler =
-        ConnectionHandlerSelect<dummy::ConnectionHandler, THandler<managed::Behaviour>>;
-    type ToSwarm = Event;
-
-    // simply delegate to underlying mDNS behaviour
-
-    delegate! {
-        to self.managed {
-            fn handle_pending_inbound_connection(&mut self, connection_id: ConnectionId, local_addr: &Multiaddr, remote_addr: &Multiaddr) -> Result<(), ConnectionDenied>;
-            fn handle_pending_outbound_connection(&mut self, connection_id: ConnectionId, maybe_peer: Option<PeerId>, addresses: &[Multiaddr], effective_role: Endpoint) -> Result<Vec<Multiaddr>, ConnectionDenied>;
-        }
-    }
-
-    fn handle_established_inbound_connection(
-        &mut self,
-        connection_id: ConnectionId,
-        peer: PeerId,
-        local_addr: &Multiaddr,
-        remote_addr: &Multiaddr,
-    ) -> Result<THandler<Self>, ConnectionDenied> {
-        Ok(ConnectionHandler::select(
-            dummy::ConnectionHandler,
-            self.managed.handle_established_inbound_connection(
-                connection_id,
-                peer,
-                local_addr,
-                remote_addr,
-            )?,
-        ))
-    }
-
-    #[allow(clippy::needless_question_mark)]
-    fn handle_established_outbound_connection(
-        &mut self,
-        connection_id: ConnectionId,
-        peer: PeerId,
-        addr: &Multiaddr,
-        role_override: Endpoint,
-        port_use: PortUse,
-    ) -> Result<THandler<Self>, ConnectionDenied> {
-        Ok(ConnectionHandler::select(
-            dummy::ConnectionHandler,
-            self.managed.handle_established_outbound_connection(
-                connection_id,
-                peer,
-                addr,
-                role_override,
-                port_use,
-            )?,
-        ))
-    }
-
-    fn on_connection_handler_event(
-        &mut self,
-        peer_id: PeerId,
-        connection_id: ConnectionId,
-        event: THandlerOutEvent<Self>,
-    ) {
-        match event {
-            Either::Left(ev) => libp2p::core::util::unreachable(ev),
-            Either::Right(ev) => {
-                self.managed
-                    .on_connection_handler_event(peer_id, connection_id, ev)
-            }
-        }
-    }
-
-    // hook into these methods to drive behavior
-
-    fn on_swarm_event(&mut self, event: FromSwarm) {
-        self.managed.on_swarm_event(event); // let mDNS handle swarm events
-
-        // handle swarm events to update internal state:
-        match event {
-            FromSwarm::ConnectionEstablished(ConnectionEstablished {
-                peer_id,
-                connection_id,
-                endpoint,
-                ..
-            }) => {
-                let remote_address = match endpoint {
-                    ConnectedPoint::Dialer { address, .. } => address,
-                    ConnectedPoint::Listener { send_back_addr, .. } => send_back_addr,
-                };
-
-                if let Some((ip, port)) = remote_address.try_to_tcp_addr() {
-                    // handle connection established event which is filtered correctly
-                    self.on_connection_established(peer_id, connection_id, ip, port)
-                }
-            }
-            FromSwarm::ConnectionClosed(ConnectionClosed {
-                peer_id,
-                connection_id,
-                endpoint,
-                ..
-            }) => {
-                let remote_address = match endpoint {
-                    ConnectedPoint::Dialer { address, .. } => address,
-                    ConnectedPoint::Listener { send_back_addr, .. } => send_back_addr,
-                };
-
-                if let Some((ip, port)) = remote_address.try_to_tcp_addr() {
-                    // handle connection closed event which is filtered correctly
-                    self.on_connection_closed(peer_id, connection_id, ip, port)
-                }
-            }
-
-            // since we are running TCP/IP transport layer, we are assuming that
-            // no address changes can occur, hence encountering one is a fatal error
-            FromSwarm::AddressChange(a) => {
-                unreachable!("unhandlable: address change encountered: {:?}", a)
-            }
-            _ => {}
-        }
-    }
-
-    fn poll(&mut self, cx: &mut Context) -> Poll<ToSwarm<Self::ToSwarm, THandlerInEvent<Self>>> {
-        // delegate to managed behaviors for any behaviors they need to perform
-        match self.managed.poll(cx) {
-            Poll::Ready(ToSwarm::GenerateEvent(e)) => {
-                match e {
-                    // handle discovered and expired events from mDNS
-                    managed::BehaviourEvent::Mdns(e) => match e.clone() {
-                        mdns::Event::Discovered(peers) => {
-                            self.handle_mdns_discovered(peers);
+impl Discovery {
+    pub async fn new(zid: ZenohId, listen_port: u16) -> io::Result<Self> {
+        let discovery_port = 52413;
+        let sock = Arc::new(UdpSocket::bind(format!("[::]:{discovery_port}")).await?);
+        //sock.set_multicast_loop_v6(false)?;
+        let ifaces: Arc<Mutex<Vec<SocketAddr>>> = Default::default();
+        let _sync = Mutex::new(
+            netwatcher::watch_interfaces_with_callback({
+                let sock = sock.clone();
+                let ifaces = ifaces.clone();
+                move |update| {
+                    for (iface_idx, iface) in update.interfaces.iter() {
+                        if iface
+                            .ipv6_ips()
+                            .all(|addr| addr.is_loopback() || addr.is_unspecified())
+                        {
+                            continue;
                         }
-                        mdns::Event::Expired(peers) => {
-                            self.handle_mdns_expired(peers);
+                        if let Err(e) = sock.join_multicast_v6(&GROUP, *iface_idx).inspect(|_| {
+                            ifaces.lock().push(SocketAddr::V6(SocketAddrV6::new(
+                                GROUP, 52413, 0, *iface_idx,
+                            )))
+                        }) {
+                            if let Some(iface) = update.interfaces.get(&iface_idx) {
+                                warn!(
+                                    "failed to join multicast v6 for interface {}: {e}",
+                                    iface.name
+                                )
+                            }
                         }
-                    },
-
-                    // handle ping events => if error then disconnect
-                    managed::BehaviourEvent::Ping(e) => {
-                        if let Err(_) = e.result {
-                            self.close_connection(e.peer, e.connection.clone())
+                    }
+                    for iface_idx in update.diff.removed {
+                        ifaces.lock().retain(|addr| {
+                            if let SocketAddr::V6(v6) = addr {
+                                v6.scope_id() != iface_idx
+                            } else {
+                                true
+                            }
+                        });
+                        if let Err(e) = sock.leave_multicast_v6(&GROUP, iface_idx) {
+                            if let Some(iface) = update.interfaces.get(&iface_idx) {
+                                warn!(
+                                    "failed to leave multicast v6 for interface {}: {e}",
+                                    iface.name
+                                )
+                            }
                         }
                     }
                 }
+            })
+            // todo: better error handling here
+            .expect("failed to bind discovery watcher"),
+        );
+        Ok(Self {
+            sock,
+            ifaces,
+            last_nonce: Default::default(),
+            listen_port,
+            zid,
+            tick: interval(Duration::from_secs(1)),
+            _sync,
+        })
+    }
 
-                // since we just consumed an event, we should immediately wake just in case
-                // there are more events to come where that came from
-                cx.waker().wake_by_ref();
-            }
-
-            // forward any other mDNS event to the swarm or its connection handler(s)
-            Poll::Ready(e) => {
-                return Poll::Ready(
-                    e.map_out(|_| unreachable!("events returning to swarm already handled"))
-                        .map_in(Either::Right),
-                );
-            }
-
-            Poll::Pending => {}
-        }
-
-        // retry connecting to all mDNS peers periodically (fails safely if already connected)
-        if self.retry_delay.poll(cx).is_ready() {
-            for (p, mas) in self.mdns_discovered.clone() {
-                for ma in mas {
-                    self.dial(p, ma)
+    pub async fn next(&mut self) -> io::Result<Discovered> {
+        let mut buf = [0u8; Hello::buf_size() + WhatsUp::buf_size() + 1];
+        loop {
+            tokio::select! {
+                _ = self.tick.tick() => {
+                    self.announce().await?;
+                }
+                res = self.sock.recv_from(&mut buf) => {
+                    let Ok((bytes_read, addr)) = res else { continue; };
+                    if let Some(discovered) = self.respond(bytes_read, addr, &mut buf).await? {
+                        return Ok(discovered)
+                    }
                 }
             }
-            // dial bootstrap peers (for environments where mDNS is unavailable)
-            for addr in &self.bootstrap_peers {
-                self.pending_events.push_back(ToSwarm::Dial {
-                    opts: DialOpts::unknown_peer_id().address(addr.clone()).build(),
-                })
+        }
+    }
+
+    async fn respond(
+        &self,
+        bytes_read: usize,
+        addr: SocketAddr,
+        buf: &mut [u8],
+    ) -> io::Result<Option<Discovered>> {
+        trace!(
+            "raw recv: {bytes_read} bytes from {addr}: {:02x?}",
+            &buf[..bytes_read]
+        );
+        if bytes_read < size_of::<Header>() {
+            trace!("dropped: early EOF");
+            return Ok(None);
+        };
+        let header: &Header = bytemuck::from_bytes(&buf[0..size_of::<Header>()]);
+        if header.magic != *b"EXO" {
+            trace!("dropped: wrong magic");
+            return Ok(None);
+        };
+        let Ok(kind) = header.kind.try_into() else {
+            trace!("dropped: unknown message kind {}", header.kind);
+            return Ok(None);
+        };
+        match kind {
+            Kind::Hello => {
+                let total = Hello::buf_size();
+                if bytes_read != total {
+                    trace!("dropped: hello wrong size");
+                    return Ok(None);
+                }
+                let hello: &Hello = bytemuck::from_bytes(&buf[size_of::<Header>()..total]);
+                if hello.nonce == *self.last_nonce.lock() {
+                    trace!("dropped: local hello nonce");
+                    return Ok(None);
+                }
+
+                // reply
+                let mut reply_buf = [0u8; WhatsUp::buf_size()];
+                let reply = WhatsUp {
+                    nonce: hello.nonce,
+                    zid: self.zid.to_le_bytes(),
+                    port_le: self.listen_port.to_le_bytes(),
+                };
+                reply.write_into(&mut reply_buf);
+
+                for i in 0..4 {
+                    if self
+                        .sock
+                        .send_to(&reply_buf, addr)
+                        .await
+                        .inspect_err(|e| debug!("send to {addr} failed: {e}"))
+                        .is_ok_and(|sent| sent == WhatsUp::buf_size())
+                    {
+                        trace!(
+                            "sent {} bytes to {addr} after {} attempt(s)",
+                            WhatsUp::buf_size(),
+                            i + 1
+                        );
+                        break;
+                    };
+                    tokio::time::sleep(Duration::from_millis(300)).await;
+                }
+                Ok(None)
+            }
+            Kind::WhatsUp => {
+                let total = WhatsUp::buf_size();
+                if bytes_read != total {
+                    trace!("dropped: whatsup wrong size");
+                    return Ok(None);
+                }
+                let whats_up: &WhatsUp = bytemuck::from_bytes(&buf[size_of::<Header>()..total]);
+                if whats_up.nonce == [0u8; 8] || whats_up.nonce != *self.last_nonce.lock() {
+                    trace!("dropped: stale nonce");
+                    return Ok(None);
+                }
+                let SocketAddr::V6(v6) = addr else {
+                    trace!("dropped: v4 addr used");
+                    return Ok(None);
+                };
+                let Ok(zid) = ZenohId::try_from(&whats_up.zid[..]) else {
+                    trace!("dropped: zenoh conversion failed");
+                    return Ok(None);
+                };
+                if zid == self.zid {
+                    trace!("dropped: self zenoh id");
+                    return Ok(None);
+                }
+                // discovered
+                let addr = {
+                    let mut x = v6.clone();
+                    x.set_port(u16::from_le_bytes(whats_up.port_le));
+                    x
+                };
+                Ok(Some(Discovered { addr, zid }))
             }
-            self.retry_delay.reset(RETRY_CONNECT_INTERVAL) // reset timeout
         }
+    }
 
-        // send out any pending events from our own service
-        if let Some(e) = self.pending_events.pop_front(cx) {
-            return Poll::Ready(e.map_in(Either::Left));
+    async fn announce(&self) -> io::Result<()> {
+        let nonce = rand::random();
+        *self.last_nonce.lock() = nonce;
+        let hello = Hello { nonce };
+
+        let mut buf = [0u8; Hello::buf_size()];
+        hello.write_into(&mut buf);
+
+        let addrs = self.ifaces.lock().clone();
+        debug!("announcing {hello:?} to {addrs:?}");
+        // rev so .remove() doesn't break things
+        for (i, addr) in addrs.into_iter().enumerate().rev() {
+            match self.sock.send_to(&buf, addr).await {
+                Ok(bytes) => trace!("sent {bytes} to {addr}"),
+                Err(e) if e.kind() == ErrorKind::HostUnreachable => {
+                    debug!("disabling discovery address {addr}: {e}");
+                    _ = self.ifaces.lock().swap_remove(i);
+                }
+                Err(e) => debug!("failed to reach {addr}: {e}"),
+            }
         }
-
-        // wait for pending events
-        Poll::Pending
+        Ok(())
     }
 }
+
+pub trait Message: Pod {
+    const KIND: Kind;
+    fn header() -> Header {
+        Header {
+            magic: *b"EXO",
+            kind: Self::KIND as u8,
+        }
+    }
+    fn write_into(&self, buf: &mut [u8]) {
+        let total = size_of::<Header>() + size_of::<Self>();
+        assert!(total <= buf.len());
+        buf[0..size_of::<Header>()].copy_from_slice(bytemuck::bytes_of(&Self::header()));
+        buf[size_of::<Header>()..total].copy_from_slice(bytemuck::bytes_of(self));
+    }
+}
+
+#[repr(u8)]
+#[derive(Debug, Clone, Copy)]
+// packet & version
+pub enum Kind {
+    Hello = 0,
+    WhatsUp = 1,
+}
+
+#[derive(Debug, Clone, Copy)]
+pub struct Discovered {
+    pub zid: ZenohId,
+    pub addr: SocketAddrV6,
+}
+
+pub struct UnknownKind;
+impl TryFrom<u8> for Kind {
+    type Error = UnknownKind;
+    fn try_from(value: u8) -> Result<Self, Self::Error> {
+        match value {
+            0 => Ok(Kind::Hello),
+            1 => Ok(Kind::WhatsUp),
+            _ => Err(UnknownKind),
+        }
+    }
+}
+#[repr(C)]
+#[derive(Debug, Clone, Copy, Pod, Zeroable)]
+pub struct Header {
+    magic: [u8; 3],
+    kind: u8,
+}
+
+#[repr(C)]
+#[derive(Debug, Clone, Copy, Pod, Zeroable)]
+pub struct Hello {
+    pub nonce: [u8; 8],
+}
+impl Hello {
+    const fn buf_size() -> usize {
+        size_of::<Header>() + size_of::<Self>()
+    }
+}
+impl Message for Hello {
+    const KIND: Kind = Kind::Hello;
+}
+
+#[repr(C)]
+#[derive(Debug, Clone, Copy, Pod, Zeroable)]
+pub struct WhatsUp {
+    pub nonce: [u8; 8],
+    pub zid: [u8; 16],
+    pub port_le: [u8; 2],
+}
+impl WhatsUp {
+    const fn buf_size() -> usize {
+        size_of::<Header>() + size_of::<Self>()
+    }
+}
+impl Message for WhatsUp {
+    const KIND: Kind = Kind::WhatsUp;
+}

rust/networking/src/lib.rs

@@ -1,44 +1,92 @@
-//! TODO: crate documentation
-//!
-//! this is here as a placeholder documentation
-//!
-//!
+use std::env;
+
+use tokio::task::JoinHandle;
+use zenoh::{Result, Session as ZSession, config::Locator};
+use zenoh_plugin_storage_manager::StoragesPlugin;
+use zenoh_plugin_trait::PluginsManager;
+
+pub use zenoh::{Config, config::ZenohId};
+
+use crate::discovery::Discovery;
+
 pub mod discovery;
 pub mod swarm;
 
-/// Namespace for all the type/trait aliases used by this crate.
-pub(crate) mod alias {
-    use std::error::Error;
-
-    pub type AnyError = Box<dyn Error + Send + Sync + 'static>;
-    pub type AnyResult<T> = Result<T, AnyError>;
+pub fn cfg(identity: u128, listen_port: u16) -> Result<zenoh::Config> {
+    assert!(listen_port != 0, "must used defined listen port port");
+    let namespace = env::var("EXO_ZENOH_NAMESPACE").unwrap_or_else(|_| "exo".to_string());
+    let mut cfg = zenoh::Config::default();
+    // todo: cleanup
+    cfg.insert_json5("id", &format!("\"{identity:x}\""))?;
+    cfg.insert_json5("mode", "\"router\"")?;
+    cfg.insert_json5("listen/endpoints", &format!("[\"tcp/[::]:{listen_port}\"]"))?;
+    cfg.insert_json5("scouting/multicast/enabled", "false")?;
+    cfg.insert_json5("scouting/multicast/autoconnect", "[]")?;
+    cfg.insert_json5("scouting/gossip/multihop", "true")?;
+    cfg.insert_json5("namespace", &format!("{namespace:?}"))?;
+    cfg.insert_json5("transport/link/tx/batch_size", "9216")?;
+    cfg.insert_json5("timestamping/enabled", "true")?;
+    cfg.insert_json5("plugins/storage_manager/__required__", "true")?;
+    cfg.insert_json5(
+        "plugins/storage_manager/storages/mem1",
+        r#"{
+            key_expr: "storage/mem1/**",
+            strip_prefix: "storage/mem1",
+            volume: "memory",
+            replication: {
+                interval: 2,
+            }
+        }"#,
+    )?;
+    Ok(cfg)
 }
 
-/// Namespace for crate-wide extension traits/methods
-pub(crate) mod ext {
-    use extend::ext;
-    use libp2p::Multiaddr;
-    use libp2p::multiaddr::Protocol;
-    use std::net::IpAddr;
+pub async fn open(cfg: zenoh::Config, listen_port: u16) -> Result<Session> {
+    assert!(listen_port != 0, "must used defined listen port");
+    let mut plugins = PluginsManager::static_plugins_only();
+    plugins.declare_static_plugin::<StoragesPlugin, _>("storage_manager", true);
+    let mut runtime = zenoh::internal::runtime::RuntimeBuilder::new(cfg)
+        .plugins_manager(plugins)
+        .build()
+        .await?;
+    let z = zenoh::session::init(runtime.clone().into()).await?;
+    runtime.start().await?;
+    let mut discovery = Discovery::new(z.zid(), listen_port).await?;
+    let _jh = tokio::task::spawn(async move {
+        loop {
+            let Ok(discovered) = discovery.next().await.inspect_err(|e| {
+                log::warn!("discovery error {e}");
+            }) else {
+                continue;
+            };
 
-    #[ext(pub, name = MultiaddrExt)]
-    impl Multiaddr {
-        /// If the multiaddress corresponds to a TCP address, extracts it
-        fn try_to_tcp_addr(&self) -> Option<(IpAddr, u16)> {
-            let mut ps = self.into_iter();
-            let ip = if let Some(p) = ps.next() {
-                match p {
-                    Protocol::Ip4(ip) => IpAddr::V4(ip),
-                    Protocol::Ip6(ip) => IpAddr::V6(ip),
-                    _ => return None,
-                }
-            } else {
-                return None;
+            if discovered.zid > runtime.zid() {
+                log::debug!("not connecting to peer with greater zid");
+                continue;
+            }
+
+            let Ok(locator) =
+                Locator::new("tcp", discovered.addr.to_string(), "").inspect_err(|e| {
+                    log::warn!("failed to pass locator from addr: {e}");
+                })
+            else {
+                continue;
             };
-            let Some(Protocol::Tcp(port)) = ps.next() else {
-                return None;
-            };
-            Some((ip, port))
+
+            runtime
+                .connect_peer(&discovered.zid.into(), &[locator])
+                .await;
         }
+    });
+    Ok(Session { z, _jh })
+}
+
+pub struct Session {
+    pub z: ZSession,
+    _jh: JoinHandle<()>,
+}
+impl Drop for Session {
+    fn drop(&mut self) {
+        self._jh.abort();
     }
 }

rust/networking/src/swarm.rs

@@ -1,24 +1,20 @@
+//! Compat shim for the old libp2p code
+
+use std::collections::HashMap;
 use std::pin::Pin;
 
-use crate::swarm::transport::tcp_transport;
-use crate::{alias, discovery};
-pub use behaviour::{Behaviour, BehaviourEvent};
-use futures_lite::{Stream, StreamExt};
-use libp2p::{PeerId, SwarmBuilder, gossipsub, identity, swarm::SwarmEvent};
-use tokio::sync::{mpsc, oneshot};
+use futures_lite::Stream;
+use tokio::sync::mpsc;
+use tokio::sync::oneshot;
+use zenoh::Result;
+use zenoh::Session;
+use zenoh::handlers::FifoChannelHandler;
+use zenoh::liveliness::LivelinessToken;
+use zenoh::pubsub::Subscriber;
+use zenoh::sample::Sample;
+use zenoh::sample::SampleKind;
 
-/// The current version of the network: this prevents devices running different versions of the
-/// software from interacting with each other.
-///
-/// TODO: right now this is a hardcoded constant; figure out what the versioning semantics should
-///       even be, and how to inject the right version into this config/initialization. E.g. should
-///       this be passed in as a parameter? What about rapidly changing versions in debug builds?
-///       this is all VERY very hard to figure out and needs to be mulled over as a team.
-pub const NETWORK_VERSION: &[u8] = b"v0.0.1";
-pub const OVERRIDE_VERSION_ENV_VAR: &str = "EXO_LIBP2P_NAMESPACE";
-
-// Uses oneshot senders to emulate function calling apis while avoiding requiring unique ownership
-// of the Swarm.
+#[derive(Debug)]
 pub enum ToSwarm {
     Unsubscribe {
         topic: String,
@@ -26,52 +22,66 @@ pub enum ToSwarm {
     },
     Subscribe {
         topic: String,
-        result_sender: oneshot::Sender<Result<bool, gossipsub::SubscriptionError>>,
+        result_sender: oneshot::Sender<Result<bool>>,
     },
     Publish {
         topic: String,
         data: Vec<u8>,
-        result_sender: oneshot::Sender<Result<gossipsub::MessageId, gossipsub::PublishError>>,
+        result_sender: oneshot::Sender<Result<()>>,
     },
 }
+#[derive(Debug)]
 pub enum FromSwarm {
-    Message {
-        from: PeerId,
-        topic: String,
-        data: Vec<u8>,
-    },
-    Discovered {
-        peer_id: PeerId,
-    },
-    Expired {
-        peer_id: PeerId,
-    },
+    Message { topic: String, data: Vec<u8> },
+    Discovered {},
+    Expired {},
 }
 
+pub type Topics = HashMap<String, Subscriber<()>>;
 pub struct Swarm {
-    swarm: libp2p::Swarm<Behaviour>,
+    pub session: crate::Session,
     from_client: mpsc::Receiver<ToSwarm>,
 }
 
 impl Swarm {
     pub fn into_stream(self) -> Pin<Box<dyn Stream<Item = FromSwarm> + Send>> {
         let Swarm {
-            mut swarm,
+            session,
             mut from_client,
         } = self;
         let stream = async_stream::stream! {
+            let mut session = session;
+            let (mut to_topics, mut from_topics) = mpsc::channel(1024);
+            let mut topics = Topics::new();
+            let Ok((_token, discovery)) = register_liveness(&mut session.z).await else { return; };
             loop {
                 tokio::select! {
                     msg = from_client.recv() => {
                         let Some(msg) = msg else { break };
-                        on_message(&mut swarm, msg);
+                        on_message(&mut session.z, &mut topics, &mut to_topics, msg).await;
                     }
-                    event = swarm.next() => {
-                        let Some(event) = event else { break };
-                        if let Some(item) = filter_swarm_event(event) {
-                            yield item;
+                    event = from_topics.recv() => {
+                        if let Some(event) = event {
+                            yield event
                         }
                     }
+                    token = discovery.recv_async() => {
+                        if let Ok(token) = token {
+                            let key_expr = token.key_expr().as_str().to_owned();
+                            let nid = key_expr.strip_prefix("nodes/").and_then(|s| s.strip_suffix("/live"));
+                            yield match token.kind() {
+                                SampleKind::Put => {
+                                    log::info!("discovered: {nid:?}");
+                                    FromSwarm::Discovered {}
+                                }
+                                SampleKind::Delete => {
+                                    log::info!("expired: {nid:?}");
+                                    FromSwarm::Expired {}
+                                }
+                            }
+                        }
+
+                    }
                 }
             }
         };
@@ -79,208 +89,100 @@ impl Swarm {
     }
 }
 
-fn on_message(swarm: &mut libp2p::Swarm<Behaviour>, message: ToSwarm) {
-    match message {
-        ToSwarm::Subscribe {
-            topic,
-            result_sender,
-        } => {
-            let result = swarm
-                .behaviour_mut()
-                .gossipsub
-                .subscribe(&gossipsub::IdentTopic::new(topic));
-            _ = result_sender.send(result);
-        }
-        ToSwarm::Unsubscribe {
-            topic,
-            result_sender,
-        } => {
-            let result = swarm
-                .behaviour_mut()
-                .gossipsub
-                .unsubscribe(&gossipsub::IdentTopic::new(topic));
-            _ = result_sender.send(result);
-        }
+async fn register_liveness(
+    session: &mut Session,
+) -> Result<(LivelinessToken, Subscriber<FifoChannelHandler<Sample>>)> {
+    let token = session
+        .liveliness()
+        .declare_token(format!("nodes/{}/live", session.zid()))
+        .await?;
+    let sub = session
+        .liveliness()
+        .declare_subscriber("nodes/*/live")
+        .history(true)
+        .await?;
+    Ok((token, sub))
+}
+
+async fn on_message(
+    session: &mut Session,
+    topics: &mut Topics,
+    to_topics: &mut mpsc::Sender<FromSwarm>,
+    msg: ToSwarm,
+) {
+    match msg {
         ToSwarm::Publish {
             topic,
             data,
             result_sender,
         } => {
-            let result = swarm
-                .behaviour_mut()
-                .gossipsub
-                .publish(gossipsub::IdentTopic::new(topic), data);
-            _ = result_sender.send(result);
+            let res = session.put(format!("topics/{topic}"), data).await;
+            _ = result_sender.send(res);
+        }
+        ToSwarm::Unsubscribe {
+            topic,
+            result_sender,
+        } => {
+            let Some((_, subscriber)) = topics.remove_entry(&topic) else {
+                _ = result_sender.send(false);
+                return;
+            };
+            _ = subscriber.undeclare().await;
+            _ = result_sender.send(true);
+        }
+        ToSwarm::Subscribe {
+            topic,
+            result_sender,
+        } => {
+            assert!(topic.is_ascii());
+            if topics.contains_key(&topic) {
+                _ = result_sender.send(Ok(false));
+                return;
+            }
+            let subscriber = match session
+                .declare_subscriber(format!("topics/{topic}"))
+                .allowed_origin(zenoh::sample::Locality::Remote)
+                .callback({
+                    let sender = to_topics.clone();
+                    let topic = topic.clone();
+                    move |sample| {
+                        if sample.kind() != SampleKind::Put {
+                            return;
+                        }
+                        _ = sender.try_send(FromSwarm::Message {
+                            topic: topic.clone(),
+                            data: sample.payload().to_bytes().to_vec(),
+                        });
+                    }
+                })
+                .await
+            {
+                Ok(p) => p,
+                Err(e) => {
+                    _ = result_sender.send(Err(e));
+                    return;
+                }
+            };
+            assert!(topics.insert(topic, subscriber).is_none());
+            _ = result_sender.send(Ok(true));
         }
     }
 }
 
-fn filter_swarm_event(event: SwarmEvent<BehaviourEvent>) -> Option<FromSwarm> {
-    match event {
-        SwarmEvent::Behaviour(BehaviourEvent::Gossipsub(gossipsub::Event::Message {
-            message:
-                gossipsub::Message {
-                    source: Some(peer_id),
-                    topic,
-                    data,
-                    ..
-                },
-            ..
-        })) => Some(FromSwarm::Message {
-            from: peer_id,
-            topic: topic.into_string(),
-            data,
-        }),
-        SwarmEvent::Behaviour(BehaviourEvent::Discovery(
-            discovery::Event::ConnectionEstablished { peer_id, .. },
-        )) => Some(FromSwarm::Discovered { peer_id }),
-        SwarmEvent::Behaviour(BehaviourEvent::Discovery(discovery::Event::ConnectionClosed {
-            peer_id,
-            ..
-        })) => Some(FromSwarm::Expired { peer_id }),
-        _ => None,
-    }
-}
-
-/// Create and configure a swarm.
-///
-/// - `listen_port`: TCP port to listen on. `0` lets the OS assign one.
-/// - `bootstrap_peers`: multiaddrs to dial for environments without mDNS.
-pub fn create_swarm(
-    keypair: identity::Keypair,
+pub async fn create_swarm(
+    identity: u128,
     from_client: mpsc::Receiver<ToSwarm>,
     bootstrap_peers: Vec<String>,
     listen_port: u16,
-) -> alias::AnyResult<Swarm> {
-    let parsed_bootstrap_peers: Vec<libp2p::Multiaddr> = bootstrap_peers
-        .iter()
-        .filter(|s| !s.is_empty())
-        .filter_map(|s| s.parse().ok())
-        .collect();
-
-    let mut swarm = SwarmBuilder::with_existing_identity(keypair)
-        .with_tokio()
-        .with_other_transport(tcp_transport)?
-        .with_behaviour(|keypair| Behaviour::new(keypair, parsed_bootstrap_peers))?
-        .build();
-
-    swarm.listen_on(format!("/ip4/0.0.0.0/tcp/{listen_port}").parse()?)?;
-    Ok(Swarm { swarm, from_client })
-}
-
-mod transport {
-    use crate::alias;
-    use crate::swarm::{NETWORK_VERSION, OVERRIDE_VERSION_ENV_VAR};
-    use futures_lite::{AsyncRead, AsyncWrite};
-    use keccak_const::Sha3_256;
-    use libp2p::core::muxing;
-    use libp2p::core::transport::Boxed;
-    use libp2p::pnet::{PnetError, PnetOutput};
-    use libp2p::{PeerId, Transport, identity, noise, pnet, yamux};
-    use std::{env, sync::LazyLock};
-
-    /// Key used for networking's private network; parametrized on the [`NETWORK_VERSION`].
-    /// See [`pnet_upgrade`] for more.
-    static PNET_PRESHARED_KEY: LazyLock<[u8; 32]> = LazyLock::new(|| {
-        let builder = Sha3_256::new().update(b"exo_discovery_network");
-
-        if let Ok(var) = env::var(OVERRIDE_VERSION_ENV_VAR) {
-            let bytes = var.into_bytes();
-            builder.update(&bytes)
-        } else {
-            builder.update(NETWORK_VERSION)
-        }
-        .finalize()
-    });
-
-    /// Make the Swarm run on a private network, as to not clash with public libp2p nodes and
-    /// also different-versioned instances of this same network.
-    /// This is implemented as an additional "upgrade" ontop of existing [`libp2p::Transport`] layers.
-    async fn pnet_upgrade<TSocket>(
-        socket: TSocket,
-        _: impl Sized,
-    ) -> Result<PnetOutput<TSocket>, PnetError>
-    where
-        TSocket: AsyncRead + AsyncWrite + Send + Unpin + 'static,
-    {
-        use pnet::{PnetConfig, PreSharedKey};
-        PnetConfig::new(PreSharedKey::new(*PNET_PRESHARED_KEY))
-            .handshake(socket)
-            .await
-    }
-
-    /// TCP/IP transport layer configuration.
-    pub fn tcp_transport(
-        keypair: &identity::Keypair,
-    ) -> alias::AnyResult<Boxed<(PeerId, muxing::StreamMuxerBox)>> {
-        use libp2p::{
-            core::upgrade::Version,
-            tcp::{Config, tokio},
-        };
-
-        // `TCP_NODELAY` enabled => avoid latency
-        let tcp_config = Config::default().nodelay(true);
-
-        // V1 + lazy flushing => 0-RTT negotiation
-        let upgrade_version = Version::V1Lazy;
-
-        // Noise is faster than TLS + we don't care much for security
-        let noise_config = noise::Config::new(keypair)?;
-
-        // Use default Yamux config for multiplexing
-        let yamux_config = yamux::Config::default();
-
-        // Create new Tokio-driven TCP/IP transport layer
-        let base_transport = tokio::Transport::new(tcp_config)
-            .and_then(pnet_upgrade)
-            .upgrade(upgrade_version)
-            .authenticate(noise_config)
-            .multiplex(yamux_config);
-
-        // Return boxed transport (to flatten complex type)
-        Ok(base_transport.boxed())
-    }
-}
-
-mod behaviour {
-    use crate::{alias, discovery};
-    use libp2p::swarm::NetworkBehaviour;
-    use libp2p::{gossipsub, identity};
-
-    /// Behavior of the Swarm which composes all desired behaviors:
-    /// Right now its just [`discovery::Behaviour`] and [`gossipsub::Behaviour`].
-    #[derive(NetworkBehaviour)]
-    pub struct Behaviour {
-        pub discovery: discovery::Behaviour,
-        pub gossipsub: gossipsub::Behaviour,
-    }
-
-    impl Behaviour {
-        pub fn new(
-            keypair: &identity::Keypair,
-            bootstrap_peers: Vec<libp2p::Multiaddr>,
-        ) -> alias::AnyResult<Self> {
-            Ok(Self {
-                discovery: discovery::Behaviour::new(keypair, bootstrap_peers)?,
-                gossipsub: gossipsub_behaviour(keypair),
-            })
-        }
-    }
-
-    fn gossipsub_behaviour(keypair: &identity::Keypair) -> gossipsub::Behaviour {
-        use gossipsub::{ConfigBuilder, MessageAuthenticity, ValidationMode};
-
-        // build a gossipsub network behaviour
-        //  => signed message authenticity + strict validation mode means the message-ID is
-        //     automatically provided by gossipsub w/out needing to provide custom message-ID function
-        gossipsub::Behaviour::new(
-            MessageAuthenticity::Signed(keypair.clone()),
-            ConfigBuilder::default()
-                .max_transmit_size(8 * 1024 * 1024)
-                .validation_mode(ValidationMode::Strict)
-                .build()
-                .expect("the configuration should always be valid"),
-        )
-        .expect("creating gossipsub behavior should always work")
-    }
+) -> Result<Swarm> {
+    // todo: bootstrap
+    if !bootstrap_peers.is_empty() || listen_port != 0 {
+        todo!();
+    }
+    let cfg = crate::cfg(identity, 52414)?;
+    let session = crate::open(cfg, 52414).await?;
+    Ok(Swarm {
+        session,
+        from_client,
+    })
 }

rust/networking/tests/bootstrap_peers.rs

@@ -1,107 +0,0 @@
-use futures_lite::StreamExt;
-use networking::swarm::{FromSwarm, create_swarm};
-use std::time::Duration;
-use tokio::sync::mpsc;
-use tokio::time::timeout;
-
-/// Helper: find a free TCP port.
-fn free_port() -> u16 {
-    let listener = std::net::TcpListener::bind("127.0.0.1:0").unwrap();
-    listener.local_addr().unwrap().port()
-}
-
-/// Two nodes connect via bootstrap peers — no mDNS needed.
-///
-/// Node A listens on a fixed port. Node B bootstraps to A's address.
-/// We verify that B emits `FromSwarm::Discovered` for A's peer ID.
-#[tokio::test]
-async fn two_nodes_connect_via_bootstrap_peers() {
-    let port_a = free_port();
-
-    // Node A: listens on a known port, no bootstrap peers
-    let keypair_a = libp2p::identity::Keypair::generate_ed25519();
-    let peer_id_a = keypair_a.public().to_peer_id();
-    let (_tx_a, rx_a) = mpsc::channel(16);
-    let swarm_a = create_swarm(keypair_a, rx_a, vec![], port_a).expect("create swarm A");
-    let mut stream_a = swarm_a.into_stream();
-
-    // Node B: bootstraps to A's address
-    let keypair_b = libp2p::identity::Keypair::generate_ed25519();
-    let (_tx_b, rx_b) = mpsc::channel(16);
-    let swarm_b = create_swarm(
-        keypair_b,
-        rx_b,
-        vec![format!("/ip4/127.0.0.1/tcp/{port_a}")],
-        0,
-    )
-    .expect("create swarm B");
-    let mut stream_b = swarm_b.into_stream();
-
-    // Wait for B to discover A (connection established)
-    let connected = timeout(Duration::from_secs(10), async {
-        loop {
-            tokio::select! {
-                Some(event) = stream_a.next() => {
-                    // A will also see B connect, but we check from B's perspective
-                    let _ = event;
-                }
-                Some(event) = stream_b.next() => {
-                    if let FromSwarm::Discovered { peer_id } = event {
-                        if peer_id == peer_id_a {
-                            return true;
-                        }
-                    }
-                }
-            }
-        }
-    })
-    .await;
-
-    assert!(
-        connected.is_ok() && connected.unwrap(),
-        "Node B should discover Node A via bootstrap peer"
-    );
-}
-
-/// Empty bootstrap peers should work (backward compatible).
-#[tokio::test]
-async fn create_swarm_with_empty_bootstrap_peers() {
-    let keypair = libp2p::identity::Keypair::generate_ed25519();
-    let (_tx, rx) = mpsc::channel(16);
-    let swarm = create_swarm(keypair, rx, vec![], 0);
-    assert!(
-        swarm.is_ok(),
-        "create_swarm with no bootstrap peers should succeed"
-    );
-}
-
-/// Invalid multiaddr strings are silently filtered out.
-#[tokio::test]
-async fn create_swarm_ignores_invalid_bootstrap_addrs() {
-    let keypair = libp2p::identity::Keypair::generate_ed25519();
-    let (_tx, rx) = mpsc::channel(16);
-    let swarm = create_swarm(
-        keypair,
-        rx,
-        vec![
-            "not-a-valid-multiaddr".to_string(),
-            "".to_string(),
-            "/ip4/10.0.0.1/tcp/30000".to_string(), // valid
-        ],
-        0,
-    );
-    assert!(
-        swarm.is_ok(),
-        "create_swarm should succeed even with invalid bootstrap addrs"
-    );
-}
-
-/// Fixed listen port works correctly.
-#[tokio::test]
-async fn create_swarm_with_fixed_port() {
-    let port = free_port();
-    let keypair = libp2p::identity::Keypair::generate_ed25519();
-    let (_tx, rx) = mpsc::channel(16);
-    let swarm = create_swarm(keypair, rx, vec![], port);
-    assert!(swarm.is_ok(), "create_swarm with fixed port should succeed");
-}

rust/networking/tests/dummy.rs

@@ -1,7 +0,0 @@
-// maybe this will hold test in the future...??
-
-#[cfg(test)]
-mod tests {
-    #[test]
-    fn does_nothing() {}
-}

rust/parts.nix

@@ -55,6 +55,7 @@
         ];
 
         OPENSSL_NO_VENDOR = "1";
+        MATURIN_NO_INSTALL_RUST = "1";
 
         # Required for pyo3 tests to find libpython
         LD_LIBRARY_PATH = lib.makeLibraryPath [ pkgs.python313 ];
@@ -81,11 +82,11 @@
       config = {
         packages = {
           # Python bindings wheel via maturin
-          exo_pyo3_bindings = craneLib.buildPackage (
+          exo-net = craneLib.buildPackage (
             commonArgs
             // {
               inherit cargoArtifacts;
-              pname = "exo_pyo3_bindings";
+              pname = "exo-net";
 
               nativeBuildInputs = commonArgs.nativeBuildInputs ++ [
                 pkgs.maturin
@@ -95,7 +96,7 @@
                 maturin build \
                   --release \
                   --manylinux off \
-                  --manifest-path rust/exo_pyo3_bindings/Cargo.toml \
+                  --manifest-path rust/exo_net/Cargo.toml \
                   --features "pyo3/extension-module,pyo3/experimental-async" \
                   --interpreter ${pkgs.python313}/bin/python \
                   --out dist

rust/util/Cargo.toml

@@ -1,15 +0,0 @@
-[package]
-name = "util"
-version = { workspace = true }
-edition = { workspace = true }
-publish = false
-
-[lib]
-doctest = false
-name = "util"
-path = "src/lib.rs"
-
-[lints]
-workspace = true
-
-[dependencies]

rust/util/src/lib.rs

@@ -1 +0,0 @@
-pub mod wakerdeque;

rust/util/src/wakerdeque.rs

@@ -1,55 +0,0 @@
-use std::collections::VecDeque;
-use std::fmt::{Debug, Formatter};
-use std::task::{Context, Waker};
-
-/// A wrapper around [`VecDeque`] which wakes (if it can) on any `push_*` methods,
-/// and updates the internally stored waker by consuming [`Context`] on any `pop_*` methods.
-pub struct WakerDeque<T> {
-    waker: Option<Waker>,
-    deque: VecDeque<T>,
-}
-
-impl<T: Debug> Debug for WakerDeque<T> {
-    fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
-        self.deque.fmt(f)
-    }
-}
-
-impl<T> WakerDeque<T> {
-    pub fn new() -> Self {
-        Self {
-            waker: None,
-            deque: VecDeque::new(),
-        }
-    }
-
-    fn update(&mut self, cx: &mut Context<'_>) {
-        self.waker = Some(cx.waker().clone());
-    }
-
-    fn wake(&mut self) {
-        let Some(ref mut w) = self.waker else { return };
-        w.wake_by_ref();
-        self.waker = None;
-    }
-
-    pub fn pop_front(&mut self, cx: &mut Context<'_>) -> Option<T> {
-        self.update(cx);
-        self.deque.pop_front()
-    }
-
-    pub fn pop_back(&mut self, cx: &mut Context<'_>) -> Option<T> {
-        self.update(cx);
-        self.deque.pop_back()
-    }
-
-    pub fn push_front(&mut self, value: T) {
-        self.wake();
-        self.deque.push_front(value);
-    }
-
-    pub fn push_back(&mut self, value: T) {
-        self.wake();
-        self.deque.push_back(value);
-    }
-}

src/exo/api/adapters/ollama.py

@@ -20,7 +20,6 @@ from exo.shared.types.chunks import (
     TokenChunk,
     ToolCallChunk,
 )
-from exo.shared.types.common import CommandId
 from exo.shared.types.text_generation import (
     Base64Image,
     InputMessage,
@@ -181,7 +180,6 @@ def ollama_request_to_text_generation(
 
 
 async def generate_ollama_chat_stream(
-    _command_id: CommandId,
     chunk_stream: AsyncGenerator[
         ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
     ],
@@ -264,7 +262,6 @@ async def generate_ollama_chat_stream(
 
 
 async def collect_ollama_chat_response(
-    _command_id: CommandId,
     chunk_stream: AsyncGenerator[
         ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
     ],
@@ -369,7 +366,6 @@ def ollama_generate_request_to_text_generation(
 
 
 async def generate_ollama_generate_stream(
-    _command_id: CommandId,
     chunk_stream: AsyncGenerator[
         ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
     ],
@@ -442,7 +438,6 @@ async def generate_ollama_generate_stream(
 
 
 async def collect_ollama_generate_response(
-    _command_id: CommandId,
     chunk_stream: AsyncGenerator[
         ErrorChunk | ToolCallChunk | TokenChunk | PrefillProgressChunk, None
     ],

src/exo/api/main.py

@@ -5,6 +5,7 @@ import json
 import random
 import time
 from collections.abc import AsyncGenerator, Awaitable, Callable, Iterable
+from dataclasses import dataclass, field
 from datetime import datetime, timezone
 from http import HTTPStatus
 from pathlib import Path
@@ -12,7 +13,7 @@ from typing import Annotated, Any, Literal, cast
 from uuid import uuid4
 
 import anyio
-from anyio import BrokenResourceError, ClosedResourceError
+from exo_net import NetSender, PySession
 from fastapi import FastAPI, File, Form, HTTPException, Query, Request, UploadFile
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import FileResponse, JSONResponse, StreamingResponse
@@ -21,6 +22,7 @@ from hypercorn.asyncio import serve  # pyright: ignore[reportUnknownVariableType
 from hypercorn.config import Config
 from hypercorn.typing import ASGIFramework
 from loguru import logger
+from pydantic import TypeAdapter
 
 from exo.api.adapters.chat_completions import (
     chat_request_to_text_generation,
@@ -140,10 +142,12 @@ from exo.shared.models.model_cards import (
 )
 from exo.shared.tracing import TraceEvent, compute_stats, export_trace, load_trace_file
 from exo.shared.types.chunks import (
-    ErrorChunk,
-    ImageChunk,
+    Chunk,
+    ImageGenerationChunk,
     InputImageChunk,
     PrefillProgressChunk,
+    StatusChunk,
+    TextGenerationChunk,
     TokenChunk,
     ToolCallChunk,
 )
@@ -157,7 +161,6 @@ from exo.shared.types.commands import (
     DeleteInstance,
     DeleteInstanceLink,
     DownloadCommand,
-    ForwarderCommand,
     ForwarderDownloadCommand,
     ImageEdits,
     ImageGeneration,
@@ -171,7 +174,6 @@ from exo.shared.types.commands import (
 )
 from exo.shared.types.common import CommandId, Id, NodeId, SystemId
 from exo.shared.types.events import (
-    ChunkGenerated,
     Event,
     IndexedEvent,
     InstanceDeleted,
@@ -230,6 +232,106 @@ def _require_disaggregation_enabled() -> None:
         )
 
 
+@dataclass
+class Transport:
+    session: PySession
+    cancel_scopes: dict[CommandId, anyio.CancelScope] = field(
+        init=False, default_factory=dict
+    )
+    command_sender: NetSender = field(init=False)
+    paused: bool = field(init=False, default=False)
+    paused_ev: anyio.Event = field(init=False, default_factory=anyio.Event)
+    tg: TaskGroup = field(init=False, default_factory=TaskGroup)
+
+    def __post_init__(self):
+        # TODO: retire root keyspace
+        self.command_sender = self.session.net_sender("orchestrator")
+
+    async def run(self):
+        async with self.tg:
+            await anyio.sleep_forever()
+
+    async def send_command(self, command: Command) -> bool:
+        while self.paused:
+            await self.paused_ev.wait()
+        return await self.command_sender.send(command.model_dump_json().encode("utf-8"))
+
+    async def stream_text(
+        self,
+        command_id: CommandId,
+    ) -> AsyncGenerator[TextGenerationChunk | StatusChunk]:
+        async for chunk in self.stream(command_id):
+            if isinstance(chunk, (TextGenerationChunk | StatusChunk)):
+                yield chunk
+
+    async def stream_images(
+        self,
+        command_id: CommandId,
+    ) -> AsyncGenerator[ImageGenerationChunk]:
+        async for chunk in self.stream(command_id):
+            if isinstance(chunk, (ImageGenerationChunk)):
+                yield chunk
+
+    async def stream(
+        self,
+        command_id: CommandId,
+    ) -> AsyncGenerator[Chunk]:
+        send, recv = channel[Chunk]()
+        self.tg.start_soon(self._run_stream, command_id, send)
+        async with recv:
+            async for item in recv:
+                yield item
+
+    async def _run_stream(self, command_id: CommandId, send: Sender[Chunk]):
+        try:
+            with anyio.CancelScope() as cs:
+                self.cancel_scopes[command_id] = cs
+                # recv from any node
+                receiver = self.session.net_receiver(
+                    f"runners/*/active_tasks/{command_id}/chunks"
+                )
+                while True:
+                    data = await receiver.recv()
+                    if data is None:
+                        logger.warning(
+                            "stream terminated early without finish reason EOF"
+                        )
+                        break
+                    await send.send(
+                        chunk := (
+                            TypeAdapter[Chunk](Chunk).validate_json(
+                                data, strict=True, extra="forbid"
+                            )
+                        )
+                    )
+                    if (
+                        not isinstance(chunk, StatusChunk)
+                        and chunk.finish_reason is not None
+                    ):
+                        break
+        except (anyio.get_cancelled_exc_class(), anyio.BrokenResourceError, anyio.ClosedResourceError):
+            with anyio.CancelScope(shield=True):
+                await self.command_sender.send(
+                    TaskCancelled(cancelled_command_id=command_id)
+                    .model_dump_json()
+                    .encode("utf-8")
+                )
+        finally:
+            self.cancel_scopes.pop(command_id, None)
+            with anyio.CancelScope(shield=True):
+                await self.command_sender.send(
+                    TaskFinished(finished_command_id=command_id)
+                    .model_dump_json()
+                    .encode("utf-8")
+                )
+
+    def cancel(self, command_id: CommandId) -> bool:
+        if (cs := self.cancel_scopes.pop(command_id, None)) is not None:
+            cs.cancel()
+            return True
+        return False
+
+
 class API:
     def __init__(
         self,
@@ -237,15 +339,14 @@ class API:
         *,
         port: int,
         event_receiver: Receiver[IndexedEvent],
-        command_sender: Sender[ForwarderCommand],
         download_command_sender: Sender[ForwarderDownloadCommand],
         # This lets us pause the API if an election is running
         election_receiver: Receiver[ElectionMessage],
+        session: PySession,
     ) -> 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.event_receiver = event_receiver
         self.election_receiver = election_receiver
@@ -254,9 +355,6 @@ class API:
         self.port = port
         self._sent_image_hashes: set[str] = set()
 
-        self.paused: bool = False
-        self.paused_ev: anyio.Event = anyio.Event()
-
         self.app = FastAPI()
 
         @self.app.middleware("http")
@@ -280,13 +378,7 @@ class API:
             name="dashboard",
         )
 
-        self._text_generation_queues: dict[
-            CommandId,
-            Sender[TokenChunk | ErrorChunk | ToolCallChunk | PrefillProgressChunk],
-        ] = {}
-        self._image_generation_queues: dict[
-            CommandId, Sender[ImageChunk | ErrorChunk]
-        ] = {}
+        self.transport = Transport(session)
         self._image_store = ImageStore(EXO_IMAGE_CACHE_DIR)
         self._tg: TaskGroup = TaskGroup()
 
@@ -307,9 +399,9 @@ class API:
     def unpause(self, result_clock: int):
         logger.info("Unpausing API")
         self.last_completed_election = result_clock
-        self.paused = False
-        self.paused_ev.set()
-        self.paused_ev = anyio.Event()
+        self.transport.paused = False
+        self.transport.paused_ev.set()
+        self.transport.paused_ev = anyio.Event()
 
     def _setup_exception_handlers(self) -> None:
         self.app.exception_handler(HTTPException)(self.http_exception_handler)
@@ -424,7 +516,7 @@ class API:
             instance_meta=payload.instance_meta,
             min_nodes=payload.min_nodes,
         )
-        await self._send(command)
+        await self.transport.send_command(command)
 
         return CreateInstanceResponse(
             message="Command received.",
@@ -449,7 +541,7 @@ class API:
         command = CreateInstance(
             instance=instance,
         )
-        await self._send(command)
+        await self.transport.send_command(command)
 
         return CreateInstanceResponse(
             message="Command received.",
@@ -632,7 +724,7 @@ class API:
         command = DeleteInstance(
             instance_id=instance_id,
         )
-        await self._send(command)
+        await self.transport.send_command(command)
         return DeleteInstanceResponse(
             message="Command received.",
             command_id=command.command_id,
@@ -667,7 +759,7 @@ class API:
             prefill_instances=list(body.prefill_instances),
             decode_instances=list(body.decode_instances),
         )
-        await self._send(command)
+        await self.transport.send_command(command)
         return InstanceLinkResponse(
             message="Command received.", command_id=command.command_id
         )
@@ -677,64 +769,24 @@ class API:
     ) -> InstanceLinkResponse:
         _require_disaggregation_enabled()
         command = DeleteInstanceLink(link_id=link_id)
-        await self._send(command)
+        await self.transport.send_command(command)
         return InstanceLinkResponse(
             message="Command received.", command_id=command.command_id
         )
 
     async def cancel_command(self, command_id: CommandId) -> CancelCommandResponse:
         """Cancel an active command by closing its stream and notifying workers."""
-        sender = self._text_generation_queues.get(
-            command_id
-        ) or self._image_generation_queues.get(command_id)
-        if sender is None:
+        if self.transport.cancel(command_id):
+            return CancelCommandResponse(
+                message="Command cancelled.",
+                command_id=command_id,
+            )
+        else:
             raise HTTPException(
                 status_code=404,
                 detail="Command not found or already completed",
             )
 
-        await self._send(TaskCancelled(cancelled_command_id=command_id))
-        sender.close()
-
-        return CancelCommandResponse(
-            message="Command cancelled.",
-            command_id=command_id,
-        )
-
-    async def _token_chunk_stream(
-        self, command_id: CommandId
-    ) -> AsyncGenerator[
-        TokenChunk | ErrorChunk | ToolCallChunk | PrefillProgressChunk, 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
-            ]()
-
-            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
-
-        except anyio.get_cancelled_exc_class():
-            command = TaskCancelled(cancelled_command_id=command_id)
-            with anyio.CancelScope(shield=True):
-                await self.command_sender.send(
-                    ForwarderCommand(origin=self._system_id, command=command)
-                )
-            raise
-        finally:
-            await self._send(TaskFinished(finished_command_id=command_id))
-            if command_id in self._text_generation_queues:
-                del self._text_generation_queues[command_id]
-
     async def _collect_text_generation_with_stats(
         self, command_id: CommandId
     ) -> BenchChatCompletionResponse:
@@ -749,7 +801,7 @@ class API:
         async with anyio.create_task_group() as tg:
             tg.start_soon(sampler.run)
 
-            async for chunk in self._token_chunk_stream(command_id):
+            async for chunk in self.transport.stream_text(command_id):
                 if isinstance(chunk, PrefillProgressChunk):
                     continue
 
@@ -816,7 +868,7 @@ class API:
         images = task_params.images
         if not images:
             command = TextGeneration(task_params=task_params)
-            await self._send(command)
+            await self.transport.send_command(command)
             return command
 
         hashes = [hashlib.sha256(img.encode("ascii")).hexdigest() for img in images]
@@ -833,7 +885,7 @@ class API:
                 new_images.append((idx, img))
 
         if not new_images:
-            await self._send(command)
+            await self.transport.send_command(command)
             return command
 
         all_chunks: list[tuple[int, str]] = []
@@ -842,7 +894,7 @@ class API:
                 all_chunks.append((img_idx, img_data[i : i + EXO_MAX_CHUNK_SIZE]))
 
         for global_idx, (img_idx, chunk_data) in enumerate(all_chunks):
-            await self._send(
+            await self.transport.send_command(
                 SendInputChunk(
                     chunk=InputImageChunk(
                         model=task_params.model,
@@ -855,7 +907,7 @@ class API:
                 )
             )
 
-        await self._send(command)
+        await self.transport.send_command(command)
         return command
 
     async def chat_completions(
@@ -875,7 +927,7 @@ class API:
                 with_sse_keepalive(
                     generate_chat_stream(
                         command.command_id,
-                        self._token_chunk_stream(command.command_id),
+                        self.transport.stream_text(command.command_id),
                     ),
                 ),
                 media_type="text/event-stream",
@@ -889,7 +941,7 @@ class API:
             return StreamingResponse(
                 collect_chat_response(
                     command.command_id,
-                    self._token_chunk_stream(command.command_id),
+                    self.transport.stream_text(command.command_id),
                 ),
                 media_type="application/json",
             )
@@ -918,7 +970,7 @@ class API:
                 with_sse_keepalive(
                     generate_chat_stream(
                         command.command_id,
-                        self._token_chunk_stream(command.command_id),
+                        self.transport.stream_text(command.command_id),
                     ),
                 ),
                 media_type="text/event-stream",
@@ -1024,7 +1076,7 @@ class API:
         command = ImageGeneration(
             task_params=payload,
         )
-        await self._send(command)
+        await self.transport.send_command(command)
 
         # Check if streaming is requested
         if payload.stream and payload.partial_images and payload.partial_images > 0:
@@ -1060,105 +1112,85 @@ class API:
         image_metadata: dict[tuple[int, bool], tuple[int | None, int | None]] = {}
         images_complete = 0
 
-        try:
-            self._image_generation_queues[command_id], recv = channel[
-                ImageChunk | ErrorChunk
-            ]()
-
-            with recv as chunks:
-                async for chunk in chunks:
-                    if chunk.finish_reason == "error":
-                        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
-
-                    key = (chunk.image_index, chunk.is_partial)
-
-                    if key not in image_chunks:
-                        image_chunks[key] = {}
-                        image_total_chunks[key] = chunk.total_chunks
-                        image_metadata[key] = (
-                            chunk.partial_index,
-                            chunk.total_partials,
-                        )
-
-                    image_chunks[key][chunk.chunk_index] = chunk.data
-
-                    # Check if this image is complete
-                    if len(image_chunks[key]) == image_total_chunks[key]:
-                        full_data = "".join(
-                            image_chunks[key][i] for i in range(len(image_chunks[key]))
-                        )
-
-                        partial_idx, total_partials = image_metadata[key]
-
-                        if chunk.is_partial:
-                            # Yield partial image event (always use b64_json for partials)
-                            event_data = {
-                                "type": "partial",
-                                "image_index": chunk.image_index,
-                                "partial_index": partial_idx,
-                                "total_partials": total_partials,
-                                "format": str(chunk.format),
-                                "data": {
-                                    "b64_json": full_data
-                                    if response_format == "b64_json"
-                                    else None,
-                                },
-                            }
-                            yield f"data: {json.dumps(event_data)}\n\n"
-                        else:
-                            # Final image
-                            if response_format == "url":
-                                image_bytes = base64.b64decode(full_data)
-                                content_type = _format_to_content_type(chunk.format)
-                                stored = self._image_store.store(
-                                    image_bytes, content_type
-                                )
-                                url = self._build_image_url(request, stored.image_id)
-                                event_data = {
-                                    "type": "final",
-                                    "image_index": chunk.image_index,
-                                    "format": str(chunk.format),
-                                    "data": {"url": url},
-                                }
-                            else:
-                                event_data = {
-                                    "type": "final",
-                                    "image_index": chunk.image_index,
-                                    "format": str(chunk.format),
-                                    "data": {"b64_json": full_data},
-                                }
-                            yield f"data: {json.dumps(event_data)}\n\n"
-                            images_complete += 1
-
-                            if images_complete >= num_images:
-                                yield "data: [DONE]\n\n"
-                                break
-
-                        # Clean up completed image chunks
-                        del image_chunks[key]
-                        del image_total_chunks[key]
-                        del image_metadata[key]
-
-        except anyio.get_cancelled_exc_class():
-            command = TaskCancelled(cancelled_command_id=command_id)
-            with anyio.CancelScope(shield=True):
-                await self.command_sender.send(
-                    ForwarderCommand(origin=self._system_id, command=command)
+        async for chunk in self.transport.stream_images(command_id):
+            if chunk.finish_reason == "error":
+                error_response = ErrorResponse(
+                    error=ErrorInfo(
+                        message=chunk.error_message or "Internal server error",
+                        type="InternalServerError",
+                        code=500,
+                    )
                 )
-            raise
-        finally:
-            await self._send(TaskFinished(finished_command_id=command_id))
-            if command_id in self._image_generation_queues:
-                del self._image_generation_queues[command_id]
+                yield f"data: {error_response.model_dump_json()}\n\n"
+                yield "data: [DONE]\n\n"
+                return
+
+            key = (chunk.image_index, chunk.is_partial)
+
+            if key not in image_chunks:
+                image_chunks[key] = {}
+                image_total_chunks[key] = chunk.total_chunks
+                image_metadata[key] = (
+                    chunk.partial_index,
+                    chunk.total_partials,
+                )
+
+            image_chunks[key][chunk.chunk_index] = chunk.data
+
+            # Check if this image is complete
+            if len(image_chunks[key]) == image_total_chunks[key]:
+                full_data = "".join(
+                    image_chunks[key][i] for i in range(len(image_chunks[key]))
+                )
+
+                partial_idx, total_partials = image_metadata[key]
+
+                if chunk.is_partial:
+                    # Yield partial image event (always use b64_json for partials)
+                    event_data = {
+                        "type": "partial",
+                        "image_index": chunk.image_index,
+                        "partial_index": partial_idx,
+                        "total_partials": total_partials,
+                        "format": str(chunk.format),
+                        "data": {
+                            "b64_json": full_data
+                            if response_format == "b64_json"
+                            else None,
+                        },
+                    }
+                    yield f"data: {json.dumps(event_data)}\n\n"
+                else:
+                    # Final image
+                    if response_format == "url":
+                        image_bytes = base64.b64decode(full_data)
+                        content_type = _format_to_content_type(chunk.format)
+                        stored = self._image_store.store(image_bytes, content_type)
+                        url = self._build_image_url(request, stored.image_id)
+                        event_data = {
+                            "type": "final",
+                            "image_index": chunk.image_index,
+                            "format": str(chunk.format),
+                            "data": {"url": url},
+                        }
+                    else:
+                        event_data = {
+                            "type": "final",
+                            "image_index": chunk.image_index,
+                            "format": str(chunk.format),
+                            "data": {"b64_json": full_data},
+                        }
+                    yield f"data: {json.dumps(event_data)}\n\n"
+                    images_complete += 1
+
+                    if images_complete >= num_images:
+                        yield "data: [DONE]\n\n"
+                        break
+
+                # Clean up completed image chunks
+                del image_chunks[key]
+                del image_total_chunks[key]
+                del image_metadata[key]
 
     async def _collect_image_chunks(
         self,
@@ -1177,74 +1209,55 @@ class API:
         images_complete = 0
         stats: ImageGenerationStats | None = None
 
-        try:
-            self._image_generation_queues[command_id], recv = channel[
-                ImageChunk | ErrorChunk
-            ]()
-
-            while images_complete < num_images:
-                with recv as chunks:
-                    async for chunk in chunks:
-                        if chunk.finish_reason == "error":
-                            raise HTTPException(
-                                status_code=500,
-                                detail=chunk.error_message or "Internal server error",
-                            )
-
-                        if chunk.is_partial:
-                            continue
-
-                        if chunk.image_index not in image_chunks:
-                            image_chunks[chunk.image_index] = {}
-                            image_total_chunks[chunk.image_index] = chunk.total_chunks
-                            image_formats[chunk.image_index] = chunk.format
-
-                        image_chunks[chunk.image_index][chunk.chunk_index] = chunk.data
-
-                        if capture_stats and chunk.stats is not None:
-                            stats = chunk.stats
-
-                        if (
-                            len(image_chunks[chunk.image_index])
-                            == image_total_chunks[chunk.image_index]
-                        ):
-                            images_complete += 1
-
-                        if images_complete >= num_images:
-                            break
-
-            images: list[ImageData] = []
-            for image_idx in range(num_images):
-                chunks_dict = image_chunks[image_idx]
-                full_data = "".join(chunks_dict[i] for i in range(len(chunks_dict)))
-                if response_format == "url" and request is not None:
-                    image_bytes = base64.b64decode(full_data)
-                    content_type = _format_to_content_type(image_formats.get(image_idx))
-                    stored = self._image_store.store(image_bytes, content_type)
-                    url = self._build_image_url(request, stored.image_id)
-                    images.append(ImageData(b64_json=None, url=url))
-                else:
-                    images.append(
-                        ImageData(
-                            b64_json=full_data
-                            if response_format == "b64_json"
-                            else None,
-                            url=None,
-                        )
+        while images_complete < num_images:
+            async for chunk in self.transport.stream_images(command_id):
+                if chunk.finish_reason == "error":
+                    raise HTTPException(
+                        status_code=500,
+                        detail=chunk.error_message or "Internal server error",
                     )
 
-            return (images, stats if capture_stats else None)
-        except anyio.get_cancelled_exc_class():
-            command = TaskCancelled(cancelled_command_id=command_id)
-            with anyio.CancelScope(shield=True):
-                await self.command_sender.send(
-                    ForwarderCommand(origin=self._system_id, command=command)
+                if chunk.is_partial:
+                    continue
+
+                if chunk.image_index not in image_chunks:
+                    image_chunks[chunk.image_index] = {}
+                    image_total_chunks[chunk.image_index] = chunk.total_chunks
+                    image_formats[chunk.image_index] = chunk.format
+
+                image_chunks[chunk.image_index][chunk.chunk_index] = chunk.data
+
+                if capture_stats and chunk.stats is not None:
+                    stats = chunk.stats
+
+                if (
+                    len(image_chunks[chunk.image_index])
+                    == image_total_chunks[chunk.image_index]
+                ):
+                    images_complete += 1
+
+                if images_complete >= num_images:
+                    break
+
+        images: list[ImageData] = []
+        for image_idx in range(num_images):
+            chunks_dict = image_chunks[image_idx]
+            full_data = "".join(chunks_dict[i] for i in range(len(chunks_dict)))
+            if response_format == "url" and request is not None:
+                image_bytes = base64.b64decode(full_data)
+                content_type = _format_to_content_type(image_formats.get(image_idx))
+                stored = self._image_store.store(image_bytes, content_type)
+                url = self._build_image_url(request, stored.image_id)
+                images.append(ImageData(b64_json=None, url=url))
+            else:
+                images.append(
+                    ImageData(
+                        b64_json=full_data if response_format == "b64_json" else None,
+                        url=None,
+                    )
                 )
-            raise
-        finally:
-            await self._send(TaskFinished(finished_command_id=command_id))
-            if command_id in self._image_generation_queues:
-                del self._image_generation_queues[command_id]
+
+        return (images, stats if capture_stats else None)
 
     async def _collect_image_generation(
         self,
@@ -1294,7 +1307,7 @@ class API:
         command = ImageGeneration(
             task_params=payload,
         )
-        await self._send(command)
+        await self.transport.send_command(command)
 
         return await self._collect_image_generation_with_stats(
             request=request,
@@ -1357,7 +1370,7 @@ class API:
             f"Sending input image: {len(image_data)} bytes in {total_chunks} chunks"
         )
         for chunk_index, chunk_data in enumerate(data_chunks):
-            await self._send(
+            await self.transport.send_command(
                 SendInputChunk(
                     chunk=InputImageChunk(
                         model=resolved_model,
@@ -1369,7 +1382,7 @@ class API:
                 )
             )
 
-        await self._send(command)
+        await self.transport.send_command(command)
         return command
 
     async def image_edits(
@@ -1497,7 +1510,7 @@ class API:
                     generate_claude_stream(
                         command.command_id,
                         payload.model,
-                        self._token_chunk_stream(command.command_id),
+                        self.transport.stream_text(command.command_id),
                     ),
                 ),
                 media_type="text/event-stream",
@@ -1512,7 +1525,7 @@ class API:
                 collect_claude_response(
                     command.command_id,
                     payload.model,
-                    self._token_chunk_stream(command.command_id),
+                    self.transport.stream_text(command.command_id),
                 ),
                 media_type="application/json",
             )
@@ -1533,7 +1546,7 @@ class API:
                     generate_responses_stream(
                         command.command_id,
                         payload.model,
-                        self._token_chunk_stream(command.command_id),
+                        self.transport.stream_text(command.command_id),
                     ),
                 ),
                 media_type="text/event-stream",
@@ -1549,7 +1562,7 @@ class API:
                 collect_responses_response(
                     command.command_id,
                     payload.model,
-                    self._token_chunk_stream(command.command_id),
+                    self.transport.stream_text(command.command_id),
                 ),
                 media_type="application/json",
             )
@@ -1575,8 +1588,7 @@ class API:
         if payload.stream:
             return StreamingResponse(
                 generate_ollama_chat_stream(
-                    command.command_id,
-                    self._token_chunk_stream(command.command_id),
+                    self.transport.stream_text(command.command_id),
                 ),
                 media_type="application/x-ndjson",
                 headers={
@@ -1588,8 +1600,7 @@ class API:
         else:
             return StreamingResponse(
                 collect_ollama_chat_response(
-                    command.command_id,
-                    self._token_chunk_stream(command.command_id),
+                    self.transport.stream_text(command.command_id),
                 ),
                 media_type="application/json",
             )
@@ -1611,8 +1622,7 @@ class API:
         if payload.stream:
             return StreamingResponse(
                 generate_ollama_generate_stream(
-                    command.command_id,
-                    self._token_chunk_stream(command.command_id),
+                    self.transport.stream_text(command.command_id),
                 ),
                 media_type="application/x-ndjson",
                 headers={
@@ -1624,8 +1634,7 @@ class API:
         else:
             return StreamingResponse(
                 collect_ollama_generate_response(
-                    command.command_id,
-                    self._token_chunk_stream(command.command_id),
+                    self.transport.stream_text(command.command_id),
                 ),
                 media_type="application/json",
             )
@@ -1761,11 +1770,8 @@ class API:
                 status_code=400, detail=f"Failed to fetch model: {exc}"
             ) from exc
 
-        await self.command_sender.send(
-            ForwarderCommand(
-                origin=self._system_id,
-                command=AddCustomModelCard(model_card=card),
-            )
+        await self.transport.command_sender.send(
+            AddCustomModelCard(model_card=card).model_dump_json().encode("utf-8")
         )
 
         # Immediately update the local cache so the subsequent GET /models
@@ -1790,11 +1796,8 @@ class API:
         if card is None or not card.is_custom:
             raise HTTPException(status_code=404, detail="Custom model card not found")
 
-        await self.command_sender.send(
-            ForwarderCommand(
-                origin=self._system_id,
-                command=DeleteCustomModelCard(model_id=model_id),
-            )
+        await self.transport.command_sender.send(
+            DeleteCustomModelCard(model_id=model_id).model_dump_json().encode("utf-8")
         )
 
         return JSONResponse(
@@ -1847,6 +1850,7 @@ class API:
         try:
             async with self._tg as tg:
                 logger.info("Starting API")
+                tg.start_soon(self.transport.run)
                 tg.start_soon(self._apply_state)
                 tg.start_soon(self._pause_on_new_election)
                 tg.start_soon(self._cleanup_expired_images)
@@ -1859,7 +1863,6 @@ class API:
                         shutdown_ev.set()
         finally:
             self._event_log.close()
-            self.command_sender.close()
             self.event_receiver.close()
 
     async def run_api(self, ev: anyio.Event):
@@ -1883,23 +1886,6 @@ class API:
                 self.state = apply(self.state, i_event)
                 event = i_event.event
 
-                if isinstance(event, ChunkGenerated):
-                    if queue := self._image_generation_queues.get(
-                        event.command_id, None
-                    ):
-                        assert isinstance(event.chunk, ImageChunk)
-                        try:
-                            await queue.send(event.chunk)
-                        except (BrokenResourceError, ClosedResourceError):
-                            self._image_generation_queues.pop(event.command_id, None)
-                    if queue := self._text_generation_queues.get(
-                        event.command_id, None
-                    ):
-                        assert not isinstance(event.chunk, ImageChunk)
-                        try:
-                            await queue.send(event.chunk)
-                        except (BrokenResourceError, ClosedResourceError):
-                            self._text_generation_queues.pop(event.command_id, None)
                 if isinstance(event, InstanceDeleted):
                     self._close_streams_for_instance(event.instance_id)
                 if isinstance(event, TracesMerged):
@@ -1914,10 +1900,7 @@ class API:
                 task, (TextGenerationTask, ImageGenerationTask, ImageEditsTask)
             ):
                 continue
-            if sender := self._text_generation_queues.pop(task.command_id, None):
-                sender.close()
-            if sender := self._image_generation_queues.pop(task.command_id, None):
-                sender.close()
+            self.transport.cancel(task.command_id)
 
     def _save_merged_trace(self, event: TracesMerged) -> None:
         traces = [
@@ -1938,7 +1921,7 @@ class API:
         with self.election_receiver as ems:
             async for message in ems:
                 if message.clock > self.last_completed_election:
-                    self.paused = True
+                    self.transport.paused = True
 
     async def _cleanup_expired_images(self):
         """Periodically clean up expired images from the store."""
@@ -1949,13 +1932,6 @@ class API:
             if removed > 0:
                 logger.debug(f"Cleaned up {removed} expired images")
 
-    async def _send(self, command: Command):
-        while self.paused:
-            await self.paused_ev.wait()
-        await self.command_sender.send(
-            ForwarderCommand(origin=self._system_id, command=command)
-        )
-
     async def _send_download(self, command: DownloadCommand):
         await self.download_command_sender.send(
             ForwarderDownloadCommand(origin=self._system_id, command=command)

src/exo/api/tests/test_cancel_command.py

@@ -1,11 +1,11 @@
 # pyright: reportUnusedFunction=false, reportAny=false
 from typing import Any
-from unittest.mock import AsyncMock, MagicMock
+from unittest.mock import AsyncMock
 
 from fastapi import FastAPI
 from fastapi.testclient import TestClient
 
-from exo.api.main import API
+from exo.api.main import API, Transport
 from exo.shared.types.common import CommandId
 
 
@@ -15,9 +15,9 @@ def _make_api() -> Any:
     app = FastAPI()
     api = object.__new__(API)
     api.app = app
-    api._text_generation_queues = {}  # pyright: ignore[reportPrivateUsage]
-    api._image_generation_queues = {}  # pyright: ignore[reportPrivateUsage]
-    api._send = AsyncMock()  # pyright: ignore[reportPrivateUsage]
+    api.transport = object.__new__(Transport)
+    api.transport.cancel = AsyncMock()
+    api.transport.send_command = AsyncMock()
     api._setup_exception_handlers()  # pyright: ignore[reportPrivateUsage]
     app.post("/v1/cancel/{command_id}")(api.cancel_command)
     return api
@@ -43,16 +43,14 @@ def test_cancel_active_text_generation() -> None:
     client = TestClient(api.app)
 
     cid = CommandId("text-cmd-123")
-    sender = MagicMock()
-    api._text_generation_queues[cid] = sender
 
     response = client.post(f"/v1/cancel/{cid}")
     assert response.status_code == 200
     data: dict[str, Any] = response.json()
     assert data["message"] == "Command cancelled."
     assert data["command_id"] == str(cid)
-    sender.close.assert_called_once()
-    api._send.assert_called_once()
+    api.transport.cancel.assert_called_once()
+    api.transport.send_command.assert_called_once()
     task_cancelled = api._send.call_args[0][0]
     assert task_cancelled.cancelled_command_id == cid
 
@@ -63,15 +61,13 @@ def test_cancel_active_image_generation() -> None:
     client = TestClient(api.app)
 
     cid = CommandId("img-cmd-456")
-    sender = MagicMock()
-    api._image_generation_queues[cid] = sender
 
     response = client.post(f"/v1/cancel/{cid}")
     assert response.status_code == 200
     data: dict[str, Any] = response.json()
     assert data["message"] == "Command cancelled."
     assert data["command_id"] == str(cid)
-    sender.close.assert_called_once()
-    api._send.assert_called_once()
-    task_cancelled = api._send.call_args[0][0]
+    api.transport.cancel.assert_called_once()
+    api.transport.send_command.assert_called_once()
+    task_cancelled = api.transport.send_command.call_args[0][0]
     assert task_cancelled.cancelled_command_id == cid

src/exo/api/tests/test_instance_deleted_stream_cleanup.py

@@ -1,9 +1,10 @@
 # pyright: reportUnusedFunction=false, reportAny=false
 """Tests that InstanceDeleted events close active generation streams."""
 
+from typing import Any
 from unittest.mock import MagicMock
 
-from exo.api.main import API
+from exo.api.main import API, Transport
 from exo.api.types import ImageGenerationTaskParams
 from exo.shared.types.common import CommandId, ModelId
 from exo.shared.types.state import State
@@ -16,12 +17,11 @@ from exo.shared.types.text_generation import (
 from exo.shared.types.worker.instances import InstanceId
 
 
-def _make_api_with_state(state: State) -> API:
+def _make_api_with_state(state: State) -> Any:
     """Create a minimal API instance with pre-set state."""
     api = object.__new__(API)
     api.state = state
-    api._text_generation_queues = {}  # pyright: ignore[reportPrivateUsage]
-    api._image_generation_queues = {}  # pyright: ignore[reportPrivateUsage]
+    api.transport = object.__new__(Transport)
     return api
 
 
@@ -47,13 +47,10 @@ def test_close_streams_for_deleted_instance() -> None:
     state = State(tasks={task.task_id: task})
     api = _make_api_with_state(state)
 
-    sender = MagicMock()
-    api._text_generation_queues[command_id] = sender  # pyright: ignore[reportPrivateUsage]
+    api._close_streams_for_instance(instance_id)
 
-    api._close_streams_for_instance(instance_id)  # pyright: ignore[reportPrivateUsage]
-
-    sender.close.assert_called_once()
-    assert command_id not in api._text_generation_queues  # pyright: ignore[reportPrivateUsage]
+    api.transport.cancel.assert_called_once()
+    assert api.transport.cancel.call_args[0][0] == command_id
 
 
 def test_close_streams_ignores_unrelated_instances() -> None:
@@ -72,7 +69,6 @@ def test_close_streams_ignores_unrelated_instances() -> None:
     api._close_streams_for_instance(target_id)  # pyright: ignore[reportPrivateUsage]
 
     sender.close.assert_not_called()
-    assert other_cmd in api._text_generation_queues  # pyright: ignore[reportPrivateUsage]
 
 
 def test_close_streams_for_deleted_instance_image_generation() -> None:

src/exo/main.py

@@ -5,8 +5,10 @@ import resource
 import signal
 from dataclasses import dataclass, field
 from typing import Self
+from uuid import uuid4
 
 import anyio
+from exo_net import PySession
 from loguru import logger
 from pydantic import PositiveInt
 
@@ -16,7 +18,7 @@ from exo.download.coordinator import DownloadCoordinator
 from exo.download.impl_shard_downloader import exo_shard_downloader
 from exo.master.main import Master
 from exo.routing.event_router import EventRouter
-from exo.routing.router import Router, get_node_id_keypair
+from exo.routing.router import Router
 from exo.shared.constants import EXO_LOG
 from exo.shared.election import Election, ElectionResult
 from exo.shared.logging import logger_cleanup, logger_setup
@@ -39,31 +41,31 @@ class Node:
     api: API | None
 
     node_id: NodeId
+    session: PySession
     offline: bool
     _api_port: int
     _tg: TaskGroup = field(init=False, default_factory=TaskGroup)
 
     @classmethod
     async def create(cls, args: "Args") -> Self:
-        keypair = get_node_id_keypair()
-        node_id = NodeId(keypair.to_node_id())
+        node_id_bytes = uuid4()
+        node_id = NodeId(str(node_id_bytes))
         session_id = SessionId(master_node_id=node_id, election_clock=0)
-        router = Router.create(
-            keypair,
+        router, session = Router.create(
+            node_id_bytes.bytes,
             bootstrap_peers=args.bootstrap_peers,
             listen_port=args.libp2p_port,
         )
         await router.register_topic(topics.GLOBAL_EVENTS)
         await router.register_topic(topics.LOCAL_EVENTS)
-        await router.register_topic(topics.COMMANDS)
         await router.register_topic(topics.ELECTION_MESSAGES)
         await router.register_topic(topics.CONNECTION_MESSAGES)
         await router.register_topic(topics.DOWNLOAD_COMMANDS)
         event_router = EventRouter(
             session_id,
-            command_sender=router.sender(topics.COMMANDS),
             external_outbound=router.sender(topics.LOCAL_EVENTS),
             external_inbound=router.receiver(topics.GLOBAL_EVENTS),
+            command_sender=session.net_sender("orchestrator"),
         )
 
         logger.info(f"Starting node {node_id}")
@@ -85,9 +87,9 @@ class Node:
                 node_id,
                 port=args.api_port,
                 event_receiver=event_router.receiver(),
-                command_sender=router.sender(topics.COMMANDS),
                 download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
                 election_receiver=router.receiver(topics.ELECTION_MESSAGES),
+                session=session,
             )
         else:
             api = None
@@ -97,9 +99,9 @@ class Node:
                 node_id,
                 event_receiver=event_router.receiver(),
                 event_sender=event_router.sender(),
-                command_sender=router.sender(topics.COMMANDS),
                 download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
                 api_port=args.api_port,
+                session=session,
             )
         else:
             worker = None
@@ -111,8 +113,8 @@ class Node:
             event_sender=event_router.sender(),
             global_event_sender=router.sender(topics.GLOBAL_EVENTS),
             local_event_receiver=router.receiver(topics.LOCAL_EVENTS),
-            command_receiver=router.receiver(topics.COMMANDS),
             download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
+            command_receiver=session.net_receiver("orchestrator"),
         )
 
         er_send, er_recv = channel[ElectionResult]()
@@ -125,7 +127,6 @@ class Node:
             election_message_sender=router.sender(topics.ELECTION_MESSAGES),
             election_message_receiver=router.receiver(topics.ELECTION_MESSAGES),
             connection_message_receiver=router.receiver(topics.CONNECTION_MESSAGES),
-            command_receiver=router.receiver(topics.COMMANDS),
             election_result_sender=er_send,
         )
 
@@ -139,6 +140,7 @@ class Node:
             master,
             api,
             node_id,
+            session,
             args.offline,
             args.api_port,
         )
@@ -188,7 +190,7 @@ class Node:
                     self.event_router.shutdown()
                     self.event_router = EventRouter(
                         result.session_id,
-                        self.router.sender(topics.COMMANDS),
+                        self.session.net_sender("orchestrator"),
                         self.router.receiver(topics.GLOBAL_EVENTS),
                         self.router.sender(topics.LOCAL_EVENTS),
                     )
@@ -209,10 +211,10 @@ class Node:
                         event_sender=self.event_router.sender(),
                         global_event_sender=self.router.sender(topics.GLOBAL_EVENTS),
                         local_event_receiver=self.router.receiver(topics.LOCAL_EVENTS),
-                        command_receiver=self.router.receiver(topics.COMMANDS),
                         download_command_sender=self.router.sender(
                             topics.DOWNLOAD_COMMANDS
                         ),
+                        command_receiver=self.session.net_receiver("orchestrator"),
                     )
                     self._tg.start_soon(self.master.run)
                 elif (
@@ -248,11 +250,11 @@ class Node:
                             self.node_id,
                             event_receiver=self.event_router.receiver(),
                             event_sender=self.event_router.sender(),
-                            command_sender=self.router.sender(topics.COMMANDS),
                             download_command_sender=self.router.sender(
                                 topics.DOWNLOAD_COMMANDS
                             ),
                             api_port=self._api_port,
+                            session=self.session,
                         )
                         self._tg.start_soon(self.worker.run)
                     if self.api:

src/exo/master/main.py

@@ -1,7 +1,9 @@
 from datetime import datetime, timedelta, timezone
 
 import anyio
+from exo_net import NetReceiver
 from loguru import logger
+from pydantic import TypeAdapter
 
 from exo.master.placement import (
     add_instance_to_placements,
@@ -15,11 +17,11 @@ from exo.shared.apply import apply
 from exo.shared.constants import EXO_EVENT_LOG_DIR, EXO_TRACING_ENABLED
 from exo.shared.types.commands import (
     AddCustomModelCard,
+    Command,
     CreateInstance,
     DeleteCustomModelCard,
     DeleteInstance,
     DeleteInstanceLink,
-    ForwarderCommand,
     ForwarderDownloadCommand,
     ImageEdits,
     ImageGeneration,
@@ -121,7 +123,7 @@ class Master:
         node_id: NodeId,
         session_id: SessionId,
         *,
-        command_receiver: Receiver[ForwarderCommand],
+        command_receiver: NetReceiver,  # todo: not this type
         event_sender: Sender[Event],
         local_event_receiver: Receiver[LocalForwarderEvent],
         global_event_sender: Sender[GlobalForwarderEvent],
@@ -155,303 +157,290 @@ class Master:
             self._event_log.close()
             self.global_event_sender.close()
             self.local_event_receiver.close()
-            self.command_receiver.close()
 
     async def shutdown(self):
         logger.info("Stopping Master")
         self._tg.cancel_tasks()
 
     async def _command_processor(self) -> None:
-        with self.command_receiver as commands:
-            async for forwarder_command in commands:
-                try:
-                    logger.info(f"Executing command: {forwarder_command.command}")
+        while True:
+            data = await self.command_receiver.recv()
+            if not data:
+                break
+            try:
+                command = TypeAdapter[Command](Command).validate_json(data)
+                logger.info(f"Executing command: {command}")
 
-                    generated_events: list[Event] = []
-                    command = forwarder_command.command
-                    instance_task_counts: dict[InstanceId, int] = {}
-                    match command:
-                        case TestCommand():
-                            pass
-                        case TextGeneration():
-                            prefill_only: set[InstanceId] = set()
-                            for link in self.state.instance_links.values():
-                                prefill_only.update(link.prefill_instances)
-                            for link in self.state.instance_links.values():
-                                prefill_only.difference_update(link.decode_instances)
+                generated_events: list[Event] = []
+                instance_task_counts: dict[InstanceId, int] = {}
+                match command:
+                    case TestCommand():
+                        pass
+                    case TextGeneration():
+                        prefill_only: set[InstanceId] = set()
+                        for link in self.state.instance_links.values():
+                            prefill_only.update(link.prefill_instances)
+                        for link in self.state.instance_links.values():
+                            prefill_only.difference_update(link.decode_instances)
 
-                            for instance in self.state.instances.values():
-                                if (
-                                    instance.shard_assignments.model_id
-                                    == command.task_params.model
-                                    and instance.instance_id not in prefill_only
-                                ):
-                                    in_flight = {TaskStatus.Pending, TaskStatus.Running}
-                                    task_count = sum(
-                                        1
-                                        for task in self.state.tasks.values()
-                                        if task.instance_id == instance.instance_id
-                                        and task.task_status in in_flight
-                                    )
-                                    instance_task_counts[instance.instance_id] = (
-                                        task_count
-                                    )
-
-                            if not instance_task_counts:
-                                raise ValueError(
-                                    f"No instance found for model {command.task_params.model}"
-                                )
-
-                            available_instance_ids = sorted(
-                                instance_task_counts.keys(),
-                                key=lambda instance_id: instance_task_counts[
-                                    instance_id
-                                ],
-                            )
-
-                            decode_instance_id = available_instance_ids[0]
-                            task_id = TaskId()
-                            params = command.task_params.model_copy(
-                                update={
-                                    "prefill_endpoint": _prefill_endpoint_for(
-                                        self.state, decode_instance_id
-                                    ),
-                                }
-                            )
-                            generated_events.append(
-                                TaskCreated(
-                                    task_id=task_id,
-                                    task=TextGenerationTask(
-                                        task_id=task_id,
-                                        command_id=command.command_id,
-                                        instance_id=decode_instance_id,
-                                        task_status=TaskStatus.Pending,
-                                        task_params=params,
-                                    ),
-                                )
-                            )
-                            self.command_task_mapping[command.command_id] = task_id
-                        case ImageGeneration():
-                            for instance in self.state.instances.values():
-                                if (
-                                    instance.shard_assignments.model_id
-                                    == command.task_params.model
-                                ):
-                                    in_flight = {TaskStatus.Pending, TaskStatus.Running}
-                                    task_count = sum(
-                                        1
-                                        for task in self.state.tasks.values()
-                                        if task.instance_id == instance.instance_id
-                                        and task.task_status in in_flight
-                                    )
-                                    instance_task_counts[instance.instance_id] = (
-                                        task_count
-                                    )
-
-                            if not instance_task_counts:
-                                raise ValueError(
-                                    f"No instance found for model {command.task_params.model}"
-                                )
-
-                            available_instance_ids = sorted(
-                                instance_task_counts.keys(),
-                                key=lambda instance_id: instance_task_counts[
-                                    instance_id
-                                ],
-                            )
-
-                            task_id = TaskId()
-                            selected_instance_id = available_instance_ids[0]
-                            generated_events.append(
-                                TaskCreated(
-                                    task_id=task_id,
-                                    task=ImageGenerationTask(
-                                        task_id=task_id,
-                                        command_id=command.command_id,
-                                        instance_id=selected_instance_id,
-                                        task_status=TaskStatus.Pending,
-                                        task_params=command.task_params,
-                                    ),
-                                )
-                            )
-
-                            self.command_task_mapping[command.command_id] = task_id
-
-                            if EXO_TRACING_ENABLED:
-                                selected_instance = self.state.instances.get(
-                                    selected_instance_id
-                                )
-                                if selected_instance:
-                                    ranks = set(
-                                        shard.device_rank
-                                        for shard in selected_instance.shard_assignments.runner_to_shard.values()
-                                    )
-                                    self._expected_ranks[task_id] = ranks
-                        case ImageEdits():
-                            for instance in self.state.instances.values():
-                                if (
-                                    instance.shard_assignments.model_id
-                                    == command.task_params.model
-                                ):
-                                    in_flight = {TaskStatus.Pending, TaskStatus.Running}
-                                    task_count = sum(
-                                        1
-                                        for task in self.state.tasks.values()
-                                        if task.instance_id == instance.instance_id
-                                        and task.task_status in in_flight
-                                    )
-                                    instance_task_counts[instance.instance_id] = (
-                                        task_count
-                                    )
-
-                            if not instance_task_counts:
-                                raise ValueError(
-                                    f"No instance found for model {command.task_params.model}"
-                                )
-
-                            available_instance_ids = sorted(
-                                instance_task_counts.keys(),
-                                key=lambda instance_id: instance_task_counts[
-                                    instance_id
-                                ],
-                            )
-
-                            task_id = TaskId()
-                            selected_instance_id = available_instance_ids[0]
-                            generated_events.append(
-                                TaskCreated(
-                                    task_id=task_id,
-                                    task=ImageEditsTask(
-                                        task_id=task_id,
-                                        command_id=command.command_id,
-                                        instance_id=selected_instance_id,
-                                        task_status=TaskStatus.Pending,
-                                        task_params=command.task_params,
-                                    ),
-                                )
-                            )
-
-                            self.command_task_mapping[command.command_id] = task_id
-
-                            if EXO_TRACING_ENABLED:
-                                selected_instance = self.state.instances.get(
-                                    selected_instance_id
-                                )
-                                if selected_instance:
-                                    ranks = set(
-                                        shard.device_rank
-                                        for shard in selected_instance.shard_assignments.runner_to_shard.values()
-                                    )
-                                    self._expected_ranks[task_id] = ranks
-                        case DeleteInstance():
-                            placement = delete_instance(command, self.state.instances)
-                            transition_events = get_transition_events(
-                                self.state.instances, placement, self.state.tasks
-                            )
-                            for cmd in cancel_unnecessary_downloads(
-                                placement, self.state.downloads
+                        for instance in self.state.instances.values():
+                            if (
+                                instance.shard_assignments.model_id
+                                == command.task_params.model
+                                and instance.instance_id not in prefill_only
                             ):
-                                await self.download_command_sender.send(
-                                    ForwarderDownloadCommand(
-                                        origin=self._system_id, command=cmd
-                                    )
-                                )
-                            generated_events.extend(transition_events)
-                        case PlaceInstance():
-                            placement = place_instance(
-                                command,
-                                self.state.topology,
-                                self.state.instances,
-                                self.state.node_memory,
-                                self.state.node_network,
-                                download_status=self.state.downloads,
-                                node_rdma_ctl=self.state.node_rdma_ctl,
-                            )
-                            transition_events = get_transition_events(
-                                self.state.instances, placement, self.state.tasks
-                            )
-                            generated_events.extend(transition_events)
-                        case CreateInstance():
-                            placement = add_instance_to_placements(
-                                command,
-                                self.state.topology,
-                                self.state.instances,
-                            )
-                            transition_events = get_transition_events(
-                                self.state.instances, placement, self.state.tasks
-                            )
-                            generated_events.extend(transition_events)
-                        case SendInputChunk(chunk=chunk):
-                            generated_events.append(
-                                InputChunkReceived(
-                                    command_id=chunk.command_id,
-                                    chunk=chunk,
-                                )
-                            )
-                        case TaskCancelled():
-                            if (
-                                task_id := self.command_task_mapping.get(
-                                    command.cancelled_command_id
-                                )
-                            ) is not None:
-                                generated_events.append(
-                                    TaskStatusUpdated(
-                                        task_status=TaskStatus.Cancelled,
-                                        task_id=task_id,
-                                    )
-                                )
-                            else:
-                                logger.warning(
-                                    f"Nonexistent command {command.cancelled_command_id} cancelled"
-                                )
-                        case TaskFinished():
-                            if (
-                                task_id := self.command_task_mapping.pop(
-                                    command.finished_command_id, None
-                                )
-                            ) is not None:
-                                generated_events.append(TaskDeleted(task_id=task_id))
-                            else:
-                                logger.warning(
-                                    f"Finished command {command.finished_command_id} finished"
+                                in_flight = {TaskStatus.Pending, TaskStatus.Running}
+                                task_count = sum(
+                                    1
+                                    for task in self.state.tasks.values()
+                                    if task.instance_id == instance.instance_id
+                                    and task.task_status in in_flight
                                 )
+                                instance_task_counts[instance.instance_id] = task_count
 
-                        case AddCustomModelCard():
-                            generated_events.append(
-                                CustomModelCardAdded(model_card=command.model_card)
+                        if not instance_task_counts:
+                            raise ValueError(
+                                f"No instance found for model {command.task_params.model}"
                             )
-                        case DeleteCustomModelCard():
-                            generated_events.append(
-                                CustomModelCardDeleted(model_id=command.model_id)
-                            )
-                        case SetInstanceLink():
-                            link = InstanceLink(
-                                link_id=command.link_id,
-                                prefill_instances=list(
-                                    dict.fromkeys(command.prefill_instances)
+
+                        available_instance_ids = sorted(
+                            instance_task_counts.keys(),
+                            key=lambda instance_id: instance_task_counts[instance_id],
+                        )
+
+                        decode_instance_id = available_instance_ids[0]
+                        task_id = TaskId()
+                        params = command.task_params.model_copy(
+                            update={
+                                "prefill_endpoint": _prefill_endpoint_for(
+                                    self.state, decode_instance_id
                                 ),
-                                decode_instances=list(
-                                    dict.fromkeys(command.decode_instances)
+                            }
+                        )
+                        generated_events.append(
+                            TaskCreated(
+                                task_id=task_id,
+                                task=TextGenerationTask(
+                                    task_id=task_id,
+                                    command_id=command.command_id,
+                                    instance_id=decode_instance_id,
+                                    task_status=TaskStatus.Pending,
+                                    task_params=params,
                                 ),
                             )
-                            generated_events.append(InstanceLinkCreated(link=link))
-                        case DeleteInstanceLink():
-                            generated_events.append(
-                                InstanceLinkDeleted(link_id=command.link_id)
-                            )
-                        case RequestEventLog():
-                            # We should just be able to send everything, since other buffers will ignore old messages
-                            # rate limit to 1000 at a time
-                            end = min(command.since_idx + 1000, len(self._event_log))
-                            for i, event in enumerate(
-                                self._event_log.read_range(command.since_idx, end),
-                                start=command.since_idx,
+                        )
+                        self.command_task_mapping[command.command_id] = task_id
+                    case ImageGeneration():
+                        for instance in self.state.instances.values():
+                            if (
+                                instance.shard_assignments.model_id
+                                == command.task_params.model
                             ):
-                                await self._send_event(IndexedEvent(idx=i, event=event))
-                    for event in generated_events:
-                        await self.event_sender.send(event)
-                except ValueError as e:
-                    logger.opt(exception=e).warning("Error in command processor")
+                                in_flight = {TaskStatus.Pending, TaskStatus.Running}
+                                task_count = sum(
+                                    1
+                                    for task in self.state.tasks.values()
+                                    if task.instance_id == instance.instance_id
+                                    and task.task_status in in_flight
+                                )
+                                instance_task_counts[instance.instance_id] = task_count
+
+                        if not instance_task_counts:
+                            raise ValueError(
+                                f"No instance found for model {command.task_params.model}"
+                            )
+
+                        available_instance_ids = sorted(
+                            instance_task_counts.keys(),
+                            key=lambda instance_id: instance_task_counts[instance_id],
+                        )
+
+                        task_id = TaskId()
+                        selected_instance_id = available_instance_ids[0]
+                        generated_events.append(
+                            TaskCreated(
+                                task_id=task_id,
+                                task=ImageGenerationTask(
+                                    task_id=task_id,
+                                    command_id=command.command_id,
+                                    instance_id=selected_instance_id,
+                                    task_status=TaskStatus.Pending,
+                                    task_params=command.task_params,
+                                ),
+                            )
+                        )
+
+                        self.command_task_mapping[command.command_id] = task_id
+
+                        if EXO_TRACING_ENABLED:
+                            selected_instance = self.state.instances.get(
+                                selected_instance_id
+                            )
+                            if selected_instance:
+                                ranks = set(
+                                    shard.device_rank
+                                    for shard in selected_instance.shard_assignments.runner_to_shard.values()
+                                )
+                                self._expected_ranks[task_id] = ranks
+                    case ImageEdits():
+                        for instance in self.state.instances.values():
+                            if (
+                                instance.shard_assignments.model_id
+                                == command.task_params.model
+                            ):
+                                in_flight = {TaskStatus.Pending, TaskStatus.Running}
+                                task_count = sum(
+                                    1
+                                    for task in self.state.tasks.values()
+                                    if task.instance_id == instance.instance_id
+                                    and task.task_status in in_flight
+                                )
+                                instance_task_counts[instance.instance_id] = task_count
+
+                        if not instance_task_counts:
+                            raise ValueError(
+                                f"No instance found for model {command.task_params.model}"
+                            )
+
+                        available_instance_ids = sorted(
+                            instance_task_counts.keys(),
+                            key=lambda instance_id: instance_task_counts[instance_id],
+                        )
+
+                        task_id = TaskId()
+                        selected_instance_id = available_instance_ids[0]
+                        generated_events.append(
+                            TaskCreated(
+                                task_id=task_id,
+                                task=ImageEditsTask(
+                                    task_id=task_id,
+                                    command_id=command.command_id,
+                                    instance_id=selected_instance_id,
+                                    task_status=TaskStatus.Pending,
+                                    task_params=command.task_params,
+                                ),
+                            )
+                        )
+
+                        self.command_task_mapping[command.command_id] = task_id
+
+                        if EXO_TRACING_ENABLED:
+                            selected_instance = self.state.instances.get(
+                                selected_instance_id
+                            )
+                            if selected_instance:
+                                ranks = set(
+                                    shard.device_rank
+                                    for shard in selected_instance.shard_assignments.runner_to_shard.values()
+                                )
+                                self._expected_ranks[task_id] = ranks
+                    case DeleteInstance():
+                        placement = delete_instance(command, self.state.instances)
+                        transition_events = get_transition_events(
+                            self.state.instances, placement, self.state.tasks
+                        )
+                        for cmd in cancel_unnecessary_downloads(
+                            placement, self.state.downloads
+                        ):
+                            await self.download_command_sender.send(
+                                ForwarderDownloadCommand(
+                                    origin=self._system_id, command=cmd
+                                )
+                            )
+                        generated_events.extend(transition_events)
+                    case PlaceInstance():
+                        placement = place_instance(
+                            command,
+                            self.state.topology,
+                            self.state.instances,
+                            self.state.node_memory,
+                            self.state.node_network,
+                            download_status=self.state.downloads,
+                            node_rdma_ctl=self.state.node_rdma_ctl,
+                        )
+                        transition_events = get_transition_events(
+                            self.state.instances, placement, self.state.tasks
+                        )
+                        generated_events.extend(transition_events)
+                    case CreateInstance():
+                        placement = add_instance_to_placements(
+                            command,
+                            self.state.topology,
+                            self.state.instances,
+                        )
+                        transition_events = get_transition_events(
+                            self.state.instances, placement, self.state.tasks
+                        )
+                        generated_events.extend(transition_events)
+                    case SendInputChunk(chunk=chunk):
+                        generated_events.append(
+                            InputChunkReceived(
+                                command_id=chunk.command_id,
+                                chunk=chunk,
+                            )
+                        )
+                    case TaskCancelled():
+                        if (
+                            task_id := self.command_task_mapping.get(
+                                command.cancelled_command_id
+                            )
+                        ) is not None:
+                            generated_events.append(
+                                TaskStatusUpdated(
+                                    task_status=TaskStatus.Cancelled,
+                                    task_id=task_id,
+                                )
+                            )
+                        else:
+                            logger.warning(
+                                f"Nonexistent command {command.cancelled_command_id} cancelled"
+                            )
+                    case TaskFinished():
+                        if (
+                            task_id := self.command_task_mapping.pop(
+                                command.finished_command_id, None
+                            )
+                        ) is not None:
+                            generated_events.append(TaskDeleted(task_id=task_id))
+                        else:
+                            logger.warning(
+                                f"Finished command {command.finished_command_id} finished"
+                            )
+
+                    case AddCustomModelCard():
+                        generated_events.append(
+                            CustomModelCardAdded(model_card=command.model_card)
+                        )
+                    case DeleteCustomModelCard():
+                        generated_events.append(
+                            CustomModelCardDeleted(model_id=command.model_id)
+                        )
+                    case SetInstanceLink():
+                        link = InstanceLink(
+                            link_id=command.link_id,
+                            prefill_instances=list(
+                                dict.fromkeys(command.prefill_instances)
+                            ),
+                            decode_instances=list(
+                                dict.fromkeys(command.decode_instances)
+                            ),
+                        )
+                        generated_events.append(InstanceLinkCreated(link=link))
+                    case DeleteInstanceLink():
+                        generated_events.append(
+                            InstanceLinkDeleted(link_id=command.link_id)
+                        )
+                    case RequestEventLog():
+                        end = len(self._event_log)
+                        for i, event in enumerate(
+                            self._event_log.read_range(command.since_idx, end),
+                            start=command.since_idx,
+                        ):
+                            await self._send_event(IndexedEvent(idx=i, event=event))
+                for event in generated_events:
+                    await self.event_sender.send(event)
+            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:

src/exo/master/tests/test_master.py

@@ -6,7 +6,6 @@ import pytest
 from loguru import logger
 
 from exo.master.main import Master
-from exo.routing.router import get_node_id_keypair
 from exo.shared.models.model_cards import ModelCard, ModelTask
 from exo.shared.types.commands import (
     CommandId,
@@ -47,8 +46,7 @@ from exo.utils.channels import channel
 
 @pytest.mark.asyncio
 async def test_master():
-    keypair = get_node_id_keypair()
-    node_id = NodeId(keypair.to_node_id())
+    node_id = NodeId("master test")
     session_id = SessionId(master_node_id=node_id, election_clock=0)
 
     ge_sender, global_event_receiver = channel[GlobalForwarderEvent]()

src/exo/routing/connection_message.py

@@ -1,15 +1,13 @@
-from exo_pyo3_bindings import PyFromSwarm
+from exo_net import PyFromSwarm
 
-from exo.shared.types.common import NodeId
 from exo.utils.pydantic_ext import FrozenModel
 
 """Serialisable types for Connection Updates/Messages"""
 
 
 class ConnectionMessage(FrozenModel):
-    node_id: NodeId
     connected: bool
 
     @classmethod
     def from_update(cls, update: PyFromSwarm.Connection) -> "ConnectionMessage":
-        return cls(node_id=NodeId(update.peer_id), connected=update.connected)
+        return cls(connected=update.connected)

src/exo/routing/event_router.py

@@ -4,9 +4,10 @@ from random import random
 import anyio
 from anyio import BrokenResourceError, ClosedResourceError
 from anyio.abc import CancelScope
+from exo_net import NetSender
 from loguru import logger
 
-from exo.shared.types.commands import ForwarderCommand, RequestEventLog
+from exo.shared.types.commands import RequestEventLog
 from exo.shared.types.common import SessionId, SystemId
 from exo.shared.types.events import (
     Event,
@@ -23,7 +24,7 @@ from exo.utils.task_group import TaskGroup
 @dataclass
 class EventRouter:
     session_id: SessionId
-    command_sender: Sender[ForwarderCommand]
+    command_sender: NetSender
     external_inbound: Receiver[GlobalForwarderEvent]
     external_outbound: Sender[LocalForwarderEvent]
     _system_id: SystemId = field(init=False, default_factory=SystemId)
@@ -152,10 +153,9 @@ class EventRouter:
                     f"Nack attempt {self._nack_attempts}: Requesting Event Log from {since_idx}"
                 )
                 await self.command_sender.send(
-                    ForwarderCommand(
-                        origin=self._system_id,
-                        command=RequestEventLog(since_idx=since_idx),
-                    )
+                    RequestEventLog(since_idx=since_idx)
+                    .model_dump_json()
+                    .encode("utf-8")
                 )
             finally:
                 if self._nack_cancel_scope is scope:

src/exo/routing/router.py

@@ -2,8 +2,6 @@ from collections.abc import Sequence
 from copy import copy
 from itertools import count
 from math import inf
-from os import PathLike
-from pathlib import Path
 from typing import cast
 
 from anyio import (
@@ -12,18 +10,9 @@ from anyio import (
     move_on_after,
     sleep_forever,
 )
-from exo_pyo3_bindings import (
-    AllQueuesFullError,
-    Keypair,
-    MessageTooLargeError,
-    NetworkingHandle,
-    NoPeersSubscribedToTopicError,
-    PyFromSwarm,
-)
-from filelock import FileLock
+from exo_net import NetworkingHandle, PyFromSwarm, PySession
 from loguru import logger
 
-from exo.shared.constants import EXO_NODE_ID_KEYPAIR
 from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.pydantic_ext import FrozenModel
 from exo.utils.task_group import TaskGroup
@@ -105,13 +94,14 @@ class Router:
     @classmethod
     def create(
         cls,
-        identity: Keypair,
+        identity: bytes,
         bootstrap_peers: Sequence[str] = (),
         listen_port: int = 0,
-    ) -> "Router":
-        return cls(
-            handle=NetworkingHandle(identity, list(bootstrap_peers), listen_port)
+    ) -> "tuple[Router, PySession]":
+        handle, session = NetworkingHandle.new(
+            identity, list(bootstrap_peers), listen_port
         )
+        return cls(handle=handle), session
 
     def __init__(self, handle: NetworkingHandle):
         self.topic_routers: dict[str, TopicRouter[FrozenModel]] = {}
@@ -191,10 +181,8 @@ class Router:
                 from_swarm = await self._net.recv()
                 logger.debug(from_swarm)
                 match from_swarm:
-                    case PyFromSwarm.Message(origin, topic, data):
-                        logger.trace(
-                            f"Received message on {topic} from {origin} with payload {data}"
-                        )
+                    case PyFromSwarm.Message(topic, data):
+                        logger.trace(f"Received message on {topic} with payload {data}")
                         if topic not in self.topic_routers:
                             logger.warning(
                                 f"Received message on unknown or inactive topic {topic}"
@@ -225,51 +213,9 @@ class Router:
     async def _networking_publish(self):
         with self.networking_receiver as networked_items:
             async for topic, data in networked_items:
-                try:
-                    logger.trace(f"Sending message on {topic} with payload {data}")
-                    if len(data) > 1024 * 1024:
-                        logger.warning(
-                            "Sending overlarge payload, network performance may be temporarily degraded"
-                        )
-                    await self._net.gossipsub_publish(topic, data)
-                except NoPeersSubscribedToTopicError:
-                    pass
-                except AllQueuesFullError:
-                    logger.warning(f"All peer queues full, dropping message on {topic}")
-                except MessageTooLargeError:
+                logger.trace(f"Sending message on {topic} with payload {data}")
+                if len(data) > 1024 * 1024:
                     logger.warning(
-                        f"Message too large for gossipsub on {topic} ({len(data)} bytes), dropping"
+                        "Sending overlarge payload, network performance may be temporarily degraded"
                     )
-
-
-def get_node_id_keypair(
-    path: str | bytes | PathLike[str] | PathLike[bytes] = EXO_NODE_ID_KEYPAIR,
-) -> Keypair:
-    """
-    Obtains the :class:`Keypair` associated with this node-ID.
-    Obtain the :class:`PeerId` by from it.
-    """
-    # TODO(evan): bring back node id persistence once we figure out how to deal with duplicates
-    return Keypair.generate()
-
-    def lock_path(path: str | bytes | PathLike[str] | PathLike[bytes]) -> Path:
-        return Path(str(path) + ".lock")
-
-    # operate with cross-process lock to avoid race conditions
-    with FileLock(lock_path(path)):
-        with open(path, "a+b") as f:  # opens in append-mode => starts at EOF
-            # if non-zero EOF, then file exists => use to get node-ID
-            if f.tell() != 0:
-                f.seek(0)  # go to start & read protobuf-encoded bytes
-                protobuf_encoded = f.read()
-
-                try:  # if decoded successfully, save & return
-                    return Keypair.from_bytes(protobuf_encoded)
-                except ValueError as e:  # on runtime error, assume corrupt file
-                    logger.warning(f"Encountered error when trying to get keypair: {e}")
-
-        # if no valid credentials, create new ones and persist
-        with open(path, "w+b") as f:
-            keypair = Keypair.generate()
-            f.write(keypair.to_bytes())
-            return keypair
+                await self._net.gossipsub_publish(topic, data)

src/exo/shared/election.py

@@ -9,7 +9,6 @@ from anyio import (
 from loguru import logger
 
 from exo.routing.connection_message import ConnectionMessage
-from exo.shared.types.commands import ForwarderCommand
 from exo.shared.types.common import NodeId, SessionId
 from exo.utils.channels import Receiver, Sender
 from exo.utils.pydantic_ext import FrozenModel
@@ -22,7 +21,6 @@ class ElectionMessage(FrozenModel):
     clock: int
     seniority: int
     proposed_session: SessionId
-    commands_seen: int
 
     # Could eventually include a list of neighbour nodes for centrality
     def __lt__(self, other: Self) -> bool:
@@ -30,8 +28,6 @@ class ElectionMessage(FrozenModel):
             return self.clock < other.clock
         if self.seniority != other.seniority:
             return self.seniority < other.seniority
-        elif self.commands_seen != other.commands_seen:
-            return self.commands_seen < other.commands_seen
         else:
             return (
                 self.proposed_session.master_node_id
@@ -54,7 +50,6 @@ class Election:
         election_message_sender: Sender[ElectionMessage],
         election_result_sender: Sender[ElectionResult],
         connection_message_receiver: Receiver[ConnectionMessage],
-        command_receiver: Receiver[ForwarderCommand],
         is_candidate: bool = True,
         seniority: int = 0,
     ):
@@ -64,7 +59,6 @@ class Election:
         self.seniority = seniority if is_candidate else -1
         self.clock = 0
         self.node_id = node_id
-        self.commands_seen = 0
         # Every node spawns as master
         self.current_session: SessionId = SessionId(
             master_node_id=node_id, election_clock=0
@@ -75,7 +69,6 @@ class Election:
         self._em_receiver = election_message_receiver
         self._er_sender = election_result_sender
         self._cm_receiver = connection_message_receiver
-        self._co_receiver = command_receiver
 
         # Campaign state
         self._candidates: list[ElectionMessage] = []
@@ -89,7 +82,6 @@ class Election:
             async with self._tg as tg:
                 tg.start_soon(self._election_receiver)
                 tg.start_soon(self._connection_receiver)
-                tg.start_soon(self._command_counter)
 
                 # And start an election immediately, that instantly resolves
                 candidates: list[ElectionMessage] = []
@@ -179,11 +171,6 @@ class Election:
                 logger.debug("Campaign started")
                 logger.debug("Connection message added")
 
-    async def _command_counter(self) -> None:
-        with self._co_receiver as commands:
-            async for _command in commands:
-                self.commands_seen += 1
-
     async def _campaign(
         self, candidates: list[ElectionMessage], campaign_timeout: float
     ) -> None:
@@ -261,5 +248,4 @@ class Election:
             ),
             clock=c,
             seniority=self.seniority,
-            commands_seen=self.commands_seen,
         )

src/exo/shared/logging.py

@@ -46,7 +46,8 @@ class _InterceptHandler(logging.Handler):
 def logger_setup(log_file: Path | None, verbosity: int = 0):
     """Set up logging for this process - formatting, file handles, verbosity and output"""
 
-    logging.getLogger("exo_pyo3_bindings").setLevel(logging.WARNING)
+    logging.getLogger("exo_net").setLevel(logging.INFO)
+    logging.getLogger("networking").setLevel(logging.INFO)
     logging.getLogger("httpx").setLevel(logging.WARNING)
     logging.getLogger("httpcore").setLevel(logging.WARNING)
 

src/exo/shared/tests/test_election.py

@@ -327,7 +327,7 @@ async def test_connection_message_triggers_new_round_broadcast() -> None:
             tg.start_soon(election.run)
 
             # Send any connection message object; we close quickly to cancel before result creation
-            await cm_tx.send(ConnectionMessage(node_id=NodeId(), connected=True))
+            await cm_tx.send(ConnectionMessage(connected=True))
 
             # Expect a broadcast for the new round at clock=1
             while True:

src/exo/shared/types/chunks.py

@@ -32,15 +32,21 @@ class TokenChunk(BaseChunk):
 
 class ErrorChunk(BaseChunk):
     error_message: str
-    finish_reason: Literal["error"] = "error"
+
+    @property
+    def finish_reason(self) -> Literal["error"]:
+        return "error"
 
 
 class ToolCallChunk(BaseChunk):
     tool_calls: list[ToolCallItem]
     usage: Usage | None
-    finish_reason: Literal["tool_calls"] = "tool_calls"
     stats: GenerationStats | None = None
 
+    @property
+    def finish_reason(self) -> Literal["tool_calls"]:
+        return "tool_calls"
+
 
 class ImageChunk(BaseChunk):
     data: str
@@ -84,7 +90,13 @@ class PrefillProgressChunk(BaseChunk):
     processed_tokens: int
     total_tokens: int
 
+    @property
+    def finish_reason(self) -> FinishReason | None:
+        return None
+
 
 StatusChunk = PrefillProgressChunk
 GenerationChunk = TokenChunk | ImageChunk | ToolCallChunk | ErrorChunk
+TextGenerationChunk = TokenChunk | ToolCallChunk | ErrorChunk
+ImageGenerationChunk = ImageChunk | ErrorChunk
 Chunk = StatusChunk | GenerationChunk

src/exo/shared/types/events.py

@@ -159,10 +159,10 @@ Event = (
     | NodeTimedOut
     | NodeGatheredInfo
     | NodeDownloadProgress
-    | ChunkGenerated
-    | InputChunkReceived
     | TopologyEdgeCreated
     | TopologyEdgeDeleted
+    | ChunkGenerated
+    | InputChunkReceived
     | TracesCollected
     | TracesMerged
     | CustomModelCardAdded

src/exo/utils/banner.py

@@ -38,7 +38,7 @@ def print_startup_banner(port: int) -> None:
 
 ╔═══════════════════════════════════════════════════════════════════════╗
 ║                                                                       ║
-║  🌐 Dashboard & API Ready                                             ║
+║  Dashboard & API Ready                                                ║
 ║                                                                       ║
 ║  {dashboard_url}{" " * (69 - len(dashboard_url))}║
 ║                                                                       ║

src/exo/worker/main.py

@@ -4,6 +4,7 @@ from datetime import datetime, timezone
 
 import anyio
 from anyio import fail_after, to_thread
+from exo_net import PySession
 from loguru import logger
 
 from exo.api.types import ImageEditsTaskParams
@@ -14,7 +15,6 @@ from exo.shared.models.model_cards import ModelId, card_cache
 from exo.shared.types.chunks import InputImageChunk
 from exo.shared.types.commands import (
     DeleteInstance,
-    ForwarderCommand,
     ForwarderDownloadCommand,
     StartDownload,
 )
@@ -65,16 +65,17 @@ class Worker:
         *,
         event_receiver: Receiver[IndexedEvent],
         event_sender: Sender[Event],
+        session: PySession,
         # This is for requesting updates. It doesn't need to be a general command sender right now,
         # but I think it's the correct way to be thinking about commands
-        command_sender: Sender[ForwarderCommand],
         download_command_sender: Sender[ForwarderDownloadCommand],
         api_port: int,
     ):
         self.node_id: NodeId = node_id
         self.event_receiver = event_receiver
         self.event_sender = event_sender
-        self.command_sender = command_sender
+        self.session = session
+        self.command_sender = session.net_sender("orchestrator")
         self.download_command_sender = download_command_sender
         self.api_port = api_port
 
@@ -114,7 +115,6 @@ class Worker:
             # Actual shutdown code - waits for all tasks to complete before executing.
             logger.info("Stopping Worker")
             self.event_sender.close()
-            self.command_sender.close()
             self.download_command_sender.close()
             for runner in self.runners.values():
                 runner.shutdown()
@@ -209,10 +209,9 @@ class Worker:
                         f"Instance {iid} exceeded {EXO_MAX_INSTANCE_RETRIES} retries, requesting deletion"
                     )
                     await self.command_sender.send(
-                        ForwarderCommand(
-                            origin=self._system_id,
-                            command=DeleteInstance(instance_id=iid),
-                        )
+                        DeleteInstance(instance_id=iid)
+                        .model_dump_json()
+                        .encode("utf-8")
                     )
                     continue
 
@@ -375,6 +374,7 @@ class Worker:
         runner = RunnerSupervisor.create(
             bound_instance=task.bound_instance,
             event_sender=self.event_sender.clone(),
+            session=self.session,
         )
         self.runners[task.bound_instance.bound_runner_id] = runner
         self._tg.start_soon(runner.run)

src/exo/worker/runner/supervisor.py

@@ -10,9 +10,11 @@ from anyio import (
     ClosedResourceError,
     to_thread,
 )
+from exo_net import NetSender, PySession
 from loguru import logger
 
 from exo.shared.types.chunks import ErrorChunk
+from exo.shared.types.commands import CommandId
 from exo.shared.types.events import (
     ChunkGenerated,
     Event,
@@ -45,20 +47,18 @@ from exo.utils.channels import MpReceiver, MpSender, Sender, mp_channel
 from exo.utils.task_group import TaskGroup
 from exo.worker.runner.bootstrap import entrypoint
 
-PREFILL_TIMEOUT_SECONDS = 60
-DECODE_TIMEOUT_SECONDS = 5
-
 
 @dataclass(eq=False)
 class RunnerSupervisor:
     shard_metadata: ShardMetadata
     bound_instance: BoundInstance
     runner_process: mp.Process
-    initialize_timeout: float
     _ev_recv: MpReceiver[Event]
     _task_sender: MpSender[Task]
     _event_sender: Sender[Event]
     _cancel_sender: MpSender[TaskId]
+    session: PySession
+
     _tg: TaskGroup = field(default_factory=TaskGroup, init=False)
     status: RunnerStatus = field(default_factory=RunnerIdle, init=False)
     pending: dict[TaskId, anyio.Event] = field(default_factory=dict, init=False)
@@ -75,7 +75,7 @@ class RunnerSupervisor:
         *,
         bound_instance: BoundInstance,
         event_sender: Sender[Event],
-        initialize_timeout: float = 400,
+        session: PySession,
     ) -> Self:
         ev_send, ev_recv = mp_channel[Event]()
         task_sender, task_recv = mp_channel[Task]()
@@ -99,11 +99,11 @@ class RunnerSupervisor:
             bound_instance=bound_instance,
             shard_metadata=shard_metadata,
             runner_process=runner_process,
-            initialize_timeout=initialize_timeout,
             _ev_recv=ev_recv,
             _task_sender=task_sender,
             _cancel_sender=cancel_sender,
             _event_sender=event_sender,
+            session=session,
         )
 
         return self
@@ -210,9 +210,26 @@ class RunnerSupervisor:
             await self._check_runner(TimeoutError("cancel pipe blocked"))
 
     async def _forward_events(self):
+        pubs: dict[CommandId, NetSender] = {}
         try:
             with self._ev_recv as events:
                 async for event in events:
+                    if isinstance(event, ChunkGenerated):
+                        if (pub := pubs.get(event.command_id, None)) is None:
+                            pub = pubs[event.command_id] = self.session.net_sender(
+                                f"runners/{self.bound_instance.bound_runner_id}/active_tasks/{event.command_id}/chunks"
+                            )
+                        sent = await pub.send(
+                            event.chunk.model_dump_json().encode("utf-8")
+                        )
+                        if not sent:
+                            logger.warning(
+                                "api node closed communication, dropping chunk"
+                            )
+                        if event.chunk.finish_reason is not None:
+                            pubs.pop(event.command_id, None)
+                        continue
+
                     if isinstance(event, RunnerStatusUpdated):
                         self.status = event.runner_status
                     if isinstance(event, TaskAcknowledged):
@@ -275,17 +292,18 @@ class RunnerSupervisor:
         for task in self.in_progress.values():
             if isinstance(task, (TextGeneration, ImageGeneration, ImageEdits)):
                 with anyio.CancelScope(shield=True):
-                    await self._event_sender.send(
-                        ChunkGenerated(
-                            command_id=task.command_id,
-                            chunk=ErrorChunk(
-                                model=self.shard_metadata.model_card.model_id,
-                                error_message=(
-                                    "Runner shutdown before completing command "
-                                    f"({cause})"
-                                ),
+                    send = self.session.net_sender(
+                        f"runners/{self.bound_instance.bound_runner_id}/active_tasks/{task.command_id}/chunks"
+                    )
+                    await send.send(
+                        ErrorChunk(
+                            model=self.shard_metadata.model_card.model_id,
+                            error_message=(
+                                f"Runner shutdown before completing command ({cause})"
                             ),
                         )
+                        .model_dump_json()
+                        .encode("utf-8")
                     )
 
         try:

tmp/state_proxy.py

@@ -0,0 +1,15 @@
+import anyio
+from exo_net import StateProxy
+
+
+async def main():
+    sp = await StateProxy.init()
+    while True:
+        data = await sp.snapshot()
+        if data != "{}":
+            print(data)
+        await anyio.sleep(1)
+
+
+if __name__ == "__main__":
+    anyio.run(main)

uv.lock

@@ -22,7 +22,7 @@ prerelease-mode = "allow"
 members = [
     "exo",
     "exo-bench",
-    "exo-pyo3-bindings",
+    "exo-net",
 ]
 constraints = [{ name = "transformers", specifier = ">=5.6.2" }]
 overrides = [
@@ -385,7 +385,7 @@ dependencies = [
     { name = "aiofiles", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "aiohttp", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "anyio", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
-    { name = "exo-pyo3-bindings", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
+    { name = "exo-net", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "fastapi", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "filelock", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "httpx", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
@@ -394,7 +394,7 @@ dependencies = [
     { name = "loguru", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "mflux", marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "mlx", version = "0.31.1", source = { registry = "https://pypi.org/simple" }, marker = "sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
-    { name = "mlx", version = "0.32.0.dev20260427+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
+    { name = "mlx", version = "0.32.0.dev20260429+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "mlx-lm", marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "mlx-vlm", marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "msgspec", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
@@ -416,21 +416,21 @@ build = [
 ]
 cpu = [
     { name = "mlx", version = "0.31.1", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
-    { name = "mlx", version = "0.32.0.dev20260427+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-cpu') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
+    { name = "mlx", version = "0.32.0.dev20260429+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-cpu') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "mlx-cpu", marker = "sys_platform == 'linux'" },
     { name = "mlx-lm", marker = "sys_platform == 'linux'" },
     { name = "mlx-vlm", marker = "sys_platform == 'linux'" },
 ]
 cuda12 = [
     { name = "mlx", version = "0.31.1", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13')" },
-    { name = "mlx", version = "0.32.0.dev20260427+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13')" },
+    { name = "mlx", version = "0.32.0.dev20260429+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13')" },
     { name = "mlx-cuda-12", marker = "sys_platform == 'linux'" },
     { name = "mlx-lm", marker = "sys_platform == 'linux'" },
     { name = "mlx-vlm", marker = "sys_platform == 'linux'" },
 ]
 cuda13 = [
     { name = "mlx", version = "0.31.1", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13')" },
-    { name = "mlx", version = "0.32.0.dev20260427+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13')" },
+    { name = "mlx", version = "0.32.0.dev20260429+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13')" },
     { name = "mlx-cuda-13", marker = "sys_platform == 'linux'" },
     { name = "mlx-lm", marker = "sys_platform == 'linux'" },
     { name = "mlx-vlm", marker = "sys_platform == 'linux'" },
@@ -451,7 +451,7 @@ requires-dist = [
     { name = "aiofiles", specifier = ">=24.1.0" },
     { name = "aiohttp", specifier = ">=3.12.14" },
     { name = "anyio", specifier = "==4.11.0" },
-    { name = "exo-pyo3-bindings", editable = "rust/exo_pyo3_bindings" },
+    { name = "exo-net", editable = "rust/exo_net" },
     { name = "fastapi", specifier = ">=0.116.1" },
     { name = "filelock", specifier = ">=3.18.0" },
     { name = "httpx", specifier = ">=0.28.1" },
@@ -541,13 +541,13 @@ requires-dist = [
 ]
 
 [[package]]
-name = "exo-pyo3-bindings"
-version = "0.2.1"
-source = { editable = "rust/exo_pyo3_bindings" }
+name = "exo-net"
+version = "0.3.0"
+source = { editable = "rust/exo_net" }
 
 [package.dev-dependencies]
 dev = [
-    { name = "exo-pyo3-bindings", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
+    { name = "exo-net", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "pytest", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "pytest-asyncio", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
 ]
@@ -556,7 +556,7 @@ dev = [
 
 [package.metadata.requires-dev]
 dev = [
-    { name = "exo-pyo3-bindings", editable = "rust/exo_pyo3_bindings" },
+    { name = "exo-net", editable = "rust/exo_net" },
     { name = "pytest", specifier = ">=8.4.0" },
     { name = "pytest-asyncio", specifier = ">=1.0.0" },
 ]
@@ -1213,7 +1213,7 @@ dependencies = [
     { name = "hf-transfer", marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "huggingface-hub", marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "matplotlib", marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
-    { name = "mlx", version = "0.32.0.dev20260427+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
+    { name = "mlx", version = "0.32.0.dev20260429+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "numpy", marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "opencv-python", marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "piexif", marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
@@ -1263,7 +1263,7 @@ wheels = [
 
 [[package]]
 name = "mlx"
-version = "0.32.0.dev20260427+cc3f3e60"
+version = "0.32.0.dev20260429+cc3f3e60"
 source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }
 resolution-markers = [
     "sys_platform == 'darwin'",
@@ -1315,7 +1315,7 @@ source = { git = "https://github.com/rltakashige/mlx-lm?branch=leo%2Fdeepseek-v4
 dependencies = [
     { name = "jinja2", marker = "sys_platform == 'darwin' or (sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "mlx", version = "0.31.1", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
-    { name = "mlx", version = "0.32.0.dev20260427+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
+    { name = "mlx", version = "0.32.0.dev20260429+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "numpy", marker = "sys_platform == 'darwin' or (sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "protobuf", marker = "sys_platform == 'darwin' or (sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "pyyaml", marker = "sys_platform == 'darwin' or (sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
@@ -1332,7 +1332,7 @@ dependencies = [
     { name = "fastapi", marker = "sys_platform == 'darwin' or (sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "miniaudio", marker = "sys_platform == 'darwin' or (sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "mlx", version = "0.31.1", source = { registry = "https://pypi.org/simple" }, marker = "(sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
-    { name = "mlx", version = "0.32.0.dev20260427+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
+    { name = "mlx", version = "0.32.0.dev20260429+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "mlx-lm", marker = "sys_platform == 'darwin' or (sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "numpy", marker = "sys_platform == 'darwin' or (sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },
     { name = "opencv-python", marker = "sys_platform == 'darwin' or (sys_platform == 'linux' and extra == 'extra-3-exo-cpu') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda12') or (sys_platform == 'linux' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda12') or (extra == 'extra-3-exo-cpu' and extra == 'extra-3-exo-cuda13') or (extra == 'extra-3-exo-cuda12' and extra == 'extra-3-exo-cuda13')" },