diff --git a/unshell-protocol/src/tests/oneshot/mod.rs b/unshell-protocol/src/tests/oneshot/mod.rs
index 93fc793..90d5657 100644
--- a/unshell-protocol/src/tests/oneshot/mod.rs
+++ b/unshell-protocol/src/tests/oneshot/mod.rs
@@ -1,3 +1,4 @@
+mod streams;
 mod support;
 
 use crate::{Endpoint, EndpointError, RouteDirection};
diff --git a/unshell-protocol/src/tests/oneshot/streams.rs b/unshell-protocol/src/tests/oneshot/streams.rs
new file mode 100644
index 0000000..31c9045
--- /dev/null
+++ b/unshell-protocol/src/tests/oneshot/streams.rs
@@ -0,0 +1,332 @@
+use crate::{Endpoint, Leaf, Packet};
+
+use alloc::{boxed::Box, format, string::ToString, vec, vec::Vec};
+
+use super::support::{CommsLeaf, ENDPOINT_A, ENDPOINT_B, assert_hook_present, assert_hook_removed};
+
+const LEAF_STREAM_CALLER: u32 = 200;
+const LEAF_STREAM_RESPONDENT: u32 = 201;
+const STREAM_HOOK_ID: u16 = 0;
+
+/// Builds the initial downwards packet that opens the stream on the respondent.
+///
+/// The request deliberately carries `end_hook = true` through `echo_packet`-style
+/// semantics: downward routing must not treat that flag as local hook cleanup. The
+/// respondent turns this into local stream state keyed by the caller's hook id.
+fn stream_open_packet(hook_id: u16) -> Packet {
+    Packet {
+        hook_id,
+        end_hook: true,
+        path: vec![ENDPOINT_A, ENDPOINT_B],
+        procedure_id: "stream.open".to_string(),
+        data: b"open".to_vec(),
+    }
+}
+
+/// Builds one upward stream frame for a previously opened hook.
+///
+/// `end_hook` is false for every intermediate frame and true only for the final
+/// frame. This is the behavior the routing layer relies on to keep hook state until
+/// the stream has actually finished sending upward.
+fn stream_frame_packet(hook_id: u16, index: usize, end_hook: bool) -> Packet {
+    Packet {
+        hook_id,
+        end_hook,
+        path: vec![ENDPOINT_A],
+        procedure_id: "stream.frame".to_string(),
+        data: format!("stream-{index}").into_bytes(),
+    }
+}
+
+/// Caller leaf that opens exactly one stream request.
+///
+/// The first allocated hook id is deterministic in these tests (`0`) because the
+/// endpoint starts with no existing hooks. Keeping the caller this small makes the
+/// per-loop stream assertions about respondent behavior rather than caller retries.
+struct StreamCallerLeaf {
+    has_run: bool,
+}
+
+/// Respondent leaf that converts the first request into a one-way stream.
+///
+/// This mimics a leaf spawning stream state, not a new endpoint: once a request is
+/// delivered locally, the leaf records the hook and emits at most one frame on each
+/// later `update`. A failed route does not advance the stream, so retry behavior can
+/// be tested by restoring the connection on a later loop.
+struct StreamRespondentLeaf {
+    stream: Option<StreamState>,
+    total_packets: usize,
+}
+
+/// In-flight stream state owned by the respondent leaf.
+///
+/// The endpoint routing layer only knows hooks and packets. This leaf-level state is
+/// the minimal application-side record needed to emit ordered frames one at a time.
+struct StreamState {
+    hook_id: u16,
+    next_index: usize,
+}
+
+impl StreamRespondentLeaf {
+    /// Creates a respondent that will emit `total_packets` stream frames.
+    fn new(total_packets: usize) -> Self {
+        Self {
+            stream: None,
+            total_packets,
+        }
+    }
+}
+
+impl Leaf for StreamCallerLeaf {
+    fn get_id(&self) -> u32 {
+        LEAF_STREAM_CALLER
+    }
+
+    fn update(&mut self, endpoint: &mut Endpoint) {
+        if self.has_run {
+            return;
+        }
+
+        let hook_id = endpoint.get_hook_id();
+        let _ = endpoint.add_outbound(stream_open_packet(hook_id));
+        self.has_run = true;
+    }
+}
+
+impl Leaf for StreamRespondentLeaf {
+    fn get_id(&self) -> u32 {
+        LEAF_STREAM_RESPONDENT
+    }
+
+    fn update(&mut self, endpoint: &mut Endpoint) {
+        self.open_stream_from_pending_request(endpoint);
+        self.send_next_frame(endpoint);
+    }
+}
+
+impl StreamRespondentLeaf {
+    /// Opens stream state from the first locally delivered request packet.
+    ///
+    /// The hook is inserted before any upward frame is routed because upward routing
+    /// is hook-gated. Additional requests are ignored while a stream is active so a
+    /// caller cannot reset ordering mid-stream in this simple one-way harness.
+    fn open_stream_from_pending_request(&mut self, endpoint: &mut Endpoint) {
+        if self.stream.is_some() {
+            return;
+        }
+
+        let local_id = endpoint.path.last().cloned().unwrap_or(0);
+        let mut opened_hook = None;
+
+        endpoint.take_inbound_clear(local_id, |packet| {
+            if opened_hook.is_none() {
+                opened_hook = Some(packet.hook_id);
+            }
+        });
+
+        if let Some(hook_id) = opened_hook {
+            endpoint.hooks.insert(hook_id, ENDPOINT_A);
+            self.stream = Some(StreamState {
+                hook_id,
+                next_index: 0,
+            });
+        }
+    }
+
+    /// Emits at most one frame for the active stream.
+    ///
+    /// The stream only advances after the routing layer accepts the packet. This is
+    /// important for final packets: a failed final route must leave hook state and
+    /// stream progress intact so the next loop can retry instead of silently losing
+    /// the end-of-stream marker.
+    fn send_next_frame(&mut self, endpoint: &mut Endpoint) {
+        let Some(stream) = self.stream.as_mut() else {
+            return;
+        };
+
+        if stream.next_index >= self.total_packets {
+            self.stream = None;
+            return;
+        }
+
+        let index = stream.next_index;
+        let end_hook = index + 1 == self.total_packets;
+        let packet = stream_frame_packet(stream.hook_id, index, end_hook);
+
+        if endpoint.add_outbound(packet).is_ok() {
+            stream.next_index += 1;
+
+            if end_hook {
+                self.stream = None;
+            }
+        }
+    }
+}
+
+/// Two endpoint, four leaf stream harness.
+///
+/// Each endpoint has exactly one application leaf and one mock connection leaf. The
+/// channel leaves are intentionally the same `CommsLeaf` used by the oneshot tests
+/// so stream behavior exercises the same serialization and routing boundary.
+fn stream_endpoints(total_packets: usize) -> (Endpoint, Endpoint) {
+    let (tx_a, rx_a) = crossbeam_channel::unbounded();
+    let (tx_b, rx_b) = crossbeam_channel::unbounded();
+
+    let mut endpoint_a = Endpoint::new(
+        ENDPOINT_A,
+        vec![
+            Box::new(StreamCallerLeaf { has_run: false }),
+            Box::new(CommsLeaf {
+                tx: tx_b,
+                rx: rx_a,
+                remote_id: ENDPOINT_B,
+                is_authority: false,
+                started: false,
+            }),
+        ],
+    );
+    endpoint_a.path = vec![ENDPOINT_A];
+
+    let mut endpoint_b = Endpoint::new(
+        ENDPOINT_B,
+        vec![
+            Box::new(StreamRespondentLeaf::new(total_packets)),
+            Box::new(CommsLeaf {
+                tx: tx_a,
+                rx: rx_b,
+                remote_id: ENDPOINT_A,
+                is_authority: true,
+                started: false,
+            }),
+        ],
+    );
+    endpoint_b.path = vec![ENDPOINT_A, ENDPOINT_B];
+
+    // Register routes before the first application packet so leaf order is not a
+    // hidden prerequisite for the initial request leaving endpoint A.
+    endpoint_a.connections.insert((ENDPOINT_B, false));
+    endpoint_b.connections.insert((ENDPOINT_A, true));
+
+    (endpoint_a, endpoint_b)
+}
+
+/// Asserts the requested two-endpoint, four-leaf topology.
+fn assert_four_leaf_topology(endpoint_a: &Endpoint, endpoint_b: &Endpoint) {
+    assert_eq!(
+        endpoint_a.leaves.len(),
+        2,
+        "caller endpoint should have two leaves"
+    );
+    assert_eq!(
+        endpoint_b.leaves.len(),
+        2,
+        "respondent endpoint should have two leaves"
+    );
+}
+
+/// Drives the initial request until it is queued locally on endpoint B.
+fn deliver_stream_request(endpoint_a: &mut Endpoint, endpoint_b: &mut Endpoint) {
+    endpoint_a.update();
+    endpoint_b.update();
+}
+
+/// Drives one respondent stream loop and delivers any produced frame to endpoint A.
+fn drive_stream_loop(endpoint_a: &mut Endpoint, endpoint_b: &mut Endpoint) {
+    endpoint_b.update();
+    endpoint_a.update();
+}
+
+/// Returns stream packets that endpoint A has received so far.
+fn received_stream_packets(endpoint: &Endpoint) -> Vec<&Packet> {
+    endpoint
+        .inbound
+        .get(&ENDPOINT_A)
+        .map(|queue| queue.iter().collect())
+        .unwrap_or_default()
+}
+
+/// Verifies ordered stream payloads and final-frame markers.
+fn assert_received_stream(endpoint: &Endpoint, expected_count: usize, final_seen: bool) {
+    let packets = received_stream_packets(endpoint);
+    assert_eq!(packets.len(), expected_count);
+
+    for (index, packet) in packets.iter().enumerate() {
+        assert_eq!(packet.hook_id, STREAM_HOOK_ID);
+        assert_eq!(packet.data, format!("stream-{index}").as_bytes());
+        assert_eq!(
+            packet.end_hook,
+            final_seen && index + 1 == expected_count,
+            "only the last received packet should close the stream"
+        );
+    }
+}
+
+#[test]
+fn one_directional_stream_returns_one_packet_per_loop() {
+    let total_packets = 3;
+    let (mut endpoint_a, mut endpoint_b) = stream_endpoints(total_packets);
+    assert_four_leaf_topology(&endpoint_a, &endpoint_b);
+
+    deliver_stream_request(&mut endpoint_a, &mut endpoint_b);
+
+    assert_received_stream(&endpoint_a, 0, false);
+    assert!(endpoint_b.hooks.is_empty());
+
+    for index in 0..total_packets {
+        drive_stream_loop(&mut endpoint_a, &mut endpoint_b);
+        let final_seen = index + 1 == total_packets;
+
+        assert_received_stream(&endpoint_a, index + 1, final_seen);
+
+        if final_seen {
+            assert_hook_removed(&endpoint_b, STREAM_HOOK_ID);
+        } else {
+            assert_hook_present(&endpoint_b, STREAM_HOOK_ID);
+        }
+    }
+}
+
+#[test]
+fn stream_does_not_emit_before_request_is_processed_by_respondent() {
+    let (mut endpoint_a, mut endpoint_b) = stream_endpoints(2);
+
+    deliver_stream_request(&mut endpoint_a, &mut endpoint_b);
+
+    assert_received_stream(&endpoint_a, 0, false);
+    assert!(endpoint_b.outbound.is_empty());
+    assert!(endpoint_b.hooks.is_empty());
+}
+
+#[test]
+fn stream_stops_after_final_packet() {
+    let total_packets = 2;
+    let (mut endpoint_a, mut endpoint_b) = stream_endpoints(total_packets);
+
+    deliver_stream_request(&mut endpoint_a, &mut endpoint_b);
+    drive_stream_loop(&mut endpoint_a, &mut endpoint_b);
+    drive_stream_loop(&mut endpoint_a, &mut endpoint_b);
+    assert_received_stream(&endpoint_a, total_packets, true);
+    assert_hook_removed(&endpoint_b, STREAM_HOOK_ID);
+
+    drive_stream_loop(&mut endpoint_a, &mut endpoint_b);
+    assert_received_stream(&endpoint_a, total_packets, true);
+    assert_hook_removed(&endpoint_b, STREAM_HOOK_ID);
+}
+
+#[test]
+fn failed_final_stream_route_keeps_hook_and_retries() {
+    let (mut endpoint_a, mut endpoint_b) = stream_endpoints(1);
+
+    deliver_stream_request(&mut endpoint_a, &mut endpoint_b);
+    endpoint_b.connections.remove(&(ENDPOINT_A, true));
+
+    drive_stream_loop(&mut endpoint_a, &mut endpoint_b);
+    assert_received_stream(&endpoint_a, 0, false);
+    assert_hook_present(&endpoint_b, STREAM_HOOK_ID);
+
+    endpoint_b.connections.insert((ENDPOINT_A, true));
+    drive_stream_loop(&mut endpoint_a, &mut endpoint_b);
+
+    assert_received_stream(&endpoint_a, 1, true);
+    assert_hook_removed(&endpoint_b, STREAM_HOOK_ID);
+}