diff --git a/Project.toml b/Project.toml
index 4fda915..25b6e82 100644
--- a/Project.toml
+++ b/Project.toml
@@ -4,7 +4,7 @@ keywords = ["Swagger", "OpenAPI", "REST"]
 license = "MIT"
 desc = "OpenAPI server and client helper for Julia"
 authors = ["JuliaHub Inc."]
-version = "0.2.6"
+version = "0.2.7"
 
 [deps]
 Base64 = "2a0f44e3-6c83-55bd-87e4-b1978d98bd5f"
@@ -35,9 +35,10 @@ p7zip_jll = "17"
 Dates = "ade2ca70-3891-5945-98fb-dc099432e06a"
 HTTP = "cd3eb016-35fb-5094-929b-558a96fad6f3"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
+Sockets = "6462fe0b-24de-5631-8697-dd941f90decc"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 TimeZones = "f269a46b-ccf7-5d73-abea-4c690281aa53"
 URIs = "5c2747f8-b7ea-4ff2-ba2e-563bfd36b1d4"
 
 [targets]
-test = ["Test", "Random", "URIs", "HTTP", "Dates", "TimeZones"]
+test = ["Test", "Random", "URIs", "HTTP", "Dates", "TimeZones", "Sockets"]
diff --git a/src/client/httplibs/juliaweb_http.jl b/src/client/httplibs/juliaweb_http.jl
index c1b4e4c..cadf0a0 100644
--- a/src/client/httplibs/juliaweb_http.jl
+++ b/src/client/httplibs/juliaweb_http.jl
@@ -317,13 +317,17 @@ function _http_streaming_request(ctx, method, url, headers, body, timeout, bytes
                     write(io, body)
                     captured_response[] = http_response = startread(io)
                     try
-                        # `read(io, n)` is unavailable on 2.0 streams; read into a reusable
-                        # buffer with `readbytes!`, which works on both 1.x and 2.x.
-                        buf = Vector{UInt8}(undef, 8192)  # 8KB chunks
+                        # `readavailable`, not `readbytes!`: on both 1.x and 2.x streams
+                        # `readbytes!(io, buf)` blocks until the whole buffer is filled
+                        # (or the body ends), so a response chunk smaller than the buffer
+                        # (e.g. a single Kubernetes watch event) is not forwarded until
+                        # enough later data accumulates — an indefinite stall on quiet
+                        # streams. `eof` blocks until data is available; `readavailable`
+                        # then returns whatever has arrived without further blocking.
                         while !eof(io)
-                            n = readbytes!(io, buf)
-                            n == 0 && break
-                            bytesread[] += write(output, view(buf, 1:n))
+                            bytes = readavailable(io)
+                            isempty(bytes) && continue
+                            bytesread[] += write(output, bytes)
                         end
                     finally
                         close(output)
diff --git a/test/runtests.jl b/test/runtests.jl
index 49f661f..6ab75ad 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -3,6 +3,7 @@ using Test, HTTP
 import OpenAPI
 
 include("chunkreader_tests.jl")
+include("streaming_latency_tests.jl")
 include("testutils.jl")
 include("modelgen/testmodelgen.jl")
 include("client/runtests.jl")
@@ -19,6 +20,9 @@ include("deep_object/deep_client.jl")
     @testset "Chunk Readers" begin
         ChunkReaderTests.runtests()
     end
+    @testset "Streaming Latency" begin
+        StreamingLatencyTests.runtests()
+    end
     @testset "Petstore Client" begin
         try
             if run_tests_with_servers && !openapi_generator_env
diff --git a/test/streaming_latency_tests.jl b/test/streaming_latency_tests.jl
new file mode 100644
index 0000000..3d366e3
--- /dev/null
+++ b/test/streaming_latency_tests.jl
@@ -0,0 +1,110 @@
+module StreamingLatencyTests
+
+using Test
+using Sockets
+
+import OpenAPI
+import OpenAPI.Clients: Client, Ctx, exec, LineChunkReader
+
+# Streamed chunks must be forwarded to the consumer as soon as they arrive,
+# not held back until an internal read buffer fills. Regression test for the
+# :http backend stalling small chunks: `readbytes!(io, buf)` blocks until the
+# whole 8KB buffer is filled, so a chunk smaller than that (e.g. a Kubernetes
+# watch event) was only delivered once enough later data accumulated — on a
+# quiet stream, effectively never.
+#
+# The server is a minimal raw-TCP chunked HTTP/1.1 responder so that the test
+# controls exactly when bytes hit the wire, independent of HTTP.jl's server
+# API (which differs between 1.x and 2.x). Two endpoints:
+#   /quick - both chunks back to back (used to warm up/compile the client path)
+#   /stall - first chunk immediately, second only after `delay` seconds
+
+# JSON string literals: in the streaming path the client parses each chunk as
+# JSON (no response object to take a content-type from), so a quoted string
+# round-trips to a `String` return value.
+const CHUNK1 = "\"tick-1\"\n"
+const CHUNK2 = "\"tick-2\"\n"
+
+function write_chunk(sock, data)
+    write(sock, string(sizeof(data), base=16), "\r\n", data, "\r\n")
+    flush(sock)
+end
+
+function handle_connection(sock, delay)
+    try
+        request_line = readline(sock)
+        path = split(request_line, ' ')[2]
+        while true
+            line = readline(sock)
+            (isempty(line) || line == "\r") && break
+        end
+        # `connection: close` so the client does not try to reuse the socket,
+        # which this bare-bones server closes after each response
+        write(sock, "HTTP/1.1 200 OK\r\ncontent-type: application/json\r\nconnection: close\r\ntransfer-encoding: chunked\r\n\r\n")
+        flush(sock)
+        write_chunk(sock, CHUNK1)
+        endswith(path, "/stall") && sleep(delay)
+        write_chunk(sock, CHUNK2)
+        write(sock, "0\r\n\r\n")
+        flush(sock)
+    catch
+        # client went away; nothing to do
+    finally
+        close(sock)
+    end
+end
+
+function start_server(delay)
+    listener = Sockets.listen(Sockets.localhost, 0)
+    _, port = Sockets.getsockname(listener)
+    @async begin
+        while true
+            sock = try
+                accept(listener)
+            catch
+                break # listener closed
+            end
+            @async handle_connection(sock, delay)
+        end
+    end
+    listener, Int(port)
+end
+
+# Run a streaming GET and return (first chunk, seconds until it arrived,
+# remaining chunks).
+function first_chunk_latency(httplib, port, path)
+    client = Client("http://127.0.0.1:$port"; httplib=httplib)
+    return_types = Dict{Regex,Type}(Regex("^200\$") => String)
+    ctx = Ctx(client, "GET", return_types, path, String[]; chunk_reader_type=LineChunkReader)
+    events = Channel{Any}(64)
+    t0 = time()
+    task = @async exec(ctx, events)
+    first = take!(events)
+    latency = time() - t0
+    remaining = collect(events)
+    wait(task)
+    (first, latency, remaining)
+end
+
+function runtests()
+    delay = 4.0
+    listener, port = start_server(delay)
+    try
+        @testset "first chunk not delayed until buffer fills ($httplib)" for httplib in values(OpenAPI.Clients.HTTPLib)
+            # warm up: compile the full streaming request path so the timing
+            # below measures I/O behavior, not JIT
+            first_chunk_latency(httplib, port, "/quick")
+
+            first, latency, remaining = first_chunk_latency(httplib, port, "/stall")
+            @test first == "tick-1"
+            # the first chunk must arrive while the server is still stalling
+            # before the second chunk, not when the response completes
+            @test latency < delay / 2
+            @test remaining == ["tick-2"]
+        end
+    finally
+        close(listener)
+    end
+end
+
+end # module StreamingLatencyTests