diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index ac8a1b9..f9b59a4 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -25,6 +25,9 @@ jobs:
           toolchain: ${{ env.RUST_VERSION }}
           components: clippy,rustfmt
 
+      - name: Cache Rust build
+        uses: Swatinem/rust-cache@v2
+
       - name: Rust format
         run: cargo fmt --check
 
@@ -51,6 +54,9 @@ jobs:
         with:
           toolchain: ${{ env.RUST_VERSION }}
 
+      - name: Cache Rust build
+        uses: Swatinem/rust-cache@v2
+
       - name: Install protoc (Linux)
         if: runner.os == 'Linux'
         run: sudo apt-get update && sudo apt-get install -y protobuf-compiler
@@ -64,4 +70,4 @@ jobs:
         run: choco install protoc -y
 
       - name: Rust integration tests against Rust binary
-        run: cargo test --locked --all-features --test cli --test formatting --test grpc --test http --test install --test network --test terminal --test update --test websocket -- --test-threads=1
+        run: cargo test --locked --all-features --test cli --test formatting --test grpc --test har --test http --test install --test network --test terminal --test update --test websocket -- --test-threads=2
diff --git a/AGENTS.md b/AGENTS.md
index b00821c..8a30113 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -30,14 +30,15 @@ Full CI-equivalent before PRs, shared transport/request/response changes, or unc
 cargo fmt --check
 cargo clippy --locked --all-targets --all-features -- -D warnings
 cargo test --locked --all-features --lib --bins
-cargo test --locked --all-features --test cli --test formatting --test grpc --test http --test install --test network --test terminal --test update --test websocket
+cargo test --locked --all-features --test cli --test formatting --test grpc --test har --test http --test install --test network --test terminal --test update --test websocket -- --test-threads=2
 ```
 
 The integration test suite uses `TcpListener::bind("127.0.0.1:0")` and
-`UdpSocket::bind("127.0.0.1:0")` for all test servers, so tests are safe to run in
-parallel. If transient "Connection refused" errors occur from port reuse races, add
-`-- --test-threads=1` to force sequential execution. The `TestServer` and
-`H3TestServer` use `mpsc` channels instead of polling for request notification.
+`UdpSocket::bind("127.0.0.1:0")` for all test servers. CI uses two test threads to
+speed up the suite without overloading timing-sensitive socket tests. If transient
+"Connection refused" or connection-reset errors occur from port reuse races, use
+`-- --test-threads=1` to diagnose sequentially. The `TestServer` and `H3TestServer`
+use `mpsc` channels instead of polling for request notification.
 
 Docs-only changes: skip Cargo unless examples/generated CLI output changed; format changed docs only:
 
diff --git a/Cargo.lock b/Cargo.lock
index 5e8f952..89c2895 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -785,6 +785,7 @@ dependencies = [
  "tokio-util",
  "tower-service",
  "url",
+ "wait-timeout",
  "webpki-roots 1.0.9",
  "windows-sys 0.61.2",
  "zip",
diff --git a/Cargo.toml b/Cargo.toml
index b635a41..1dde9b8 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -86,3 +86,4 @@ predicates = "=3.1.4"
 rcgen = { version = "=0.14.8", features = ["aws_lc_rs"] }
 sha1 = "=0.11.0"
 tempfile = "=3.27.0"
+wait-timeout = "=0.2.1"
diff --git a/tests/support/common.rs b/tests/support/common.rs
index 935e495..9db1a23 100644
--- a/tests/support/common.rs
+++ b/tests/support/common.rs
@@ -10,6 +10,7 @@ use std::time::{Duration, Instant};
 
 use tempfile::TempDir;
 use url::Url;
+use wait_timeout::ChildExt;
 
 #[cfg(unix)]
 use std::os::fd::{FromRawFd, RawFd};
@@ -113,30 +114,41 @@ pub(crate) fn run_fetch_once(opts: FetchOpts, args: &[&str]) -> FetchOutput {
         let mut stdin = child.stdin.take().expect("child stdin");
         stdin.write_all(input.as_bytes()).expect("write stdin");
     }
-    let start = Instant::now();
-    loop {
-        if child.try_wait().expect("poll fetch").is_some() {
-            break;
-        }
-        if start.elapsed() > Duration::from_secs(15) {
-            let _ = child.kill();
-            let out = child.wait_with_output().expect("wait killed fetch");
-            return FetchOutput {
-                status: out.status,
-                stdout: String::from_utf8_lossy(&out.stdout).into_owned(),
-                stderr: format!(
-                    "{}\nfetch test harness timeout after 30s",
-                    String::from_utf8_lossy(&out.stderr)
-                ),
-            };
-        }
-        thread::sleep(Duration::from_millis(10));
+
+    // Drain both pipes while the child runs so a verbose command cannot block on
+    // a full pipe. wait_timeout avoids adding up to 10ms of polling latency to
+    // every short-lived CLI invocation in the integration suite.
+    let mut stdout = child.stdout.take().expect("child stdout");
+    let stdout_reader = thread::spawn(move || {
+        let mut output = Vec::new();
+        stdout.read_to_end(&mut output).expect("read fetch stdout");
+        output
+    });
+    let mut stderr = child.stderr.take().expect("child stderr");
+    let stderr_reader = thread::spawn(move || {
+        let mut output = Vec::new();
+        stderr.read_to_end(&mut output).expect("read fetch stderr");
+        output
+    });
+
+    let timed_out = child
+        .wait_timeout(Duration::from_secs(15))
+        .expect("wait for fetch")
+        .is_none();
+    if timed_out {
+        let _ = child.kill();
+    }
+    let status = child.wait().expect("wait fetch");
+    let stdout = stdout_reader.join().expect("join fetch stdout reader");
+    let stderr = stderr_reader.join().expect("join fetch stderr reader");
+    let mut stderr = String::from_utf8_lossy(&stderr).into_owned();
+    if timed_out {
+        stderr.push_str("\nfetch test harness timeout after 15s");
     }
-    let out = child.wait_with_output().expect("wait fetch");
     FetchOutput {
-        status: out.status,
-        stdout: String::from_utf8_lossy(&out.stdout).into_owned(),
-        stderr: String::from_utf8_lossy(&out.stderr).into_owned(),
+        status,
+        stdout: String::from_utf8_lossy(&stdout).into_owned(),
+        stderr,
     }
 }
 
@@ -286,17 +298,10 @@ pub(crate) fn wait_child(
     child: &mut std::process::Child,
     timeout: Duration,
 ) -> Option<std::io::Result<ExitStatus>> {
-    let start = Instant::now();
-    loop {
-        match child.try_wait() {
-            Ok(Some(status)) => return Some(Ok(status)),
-            Ok(None) => {}
-            Err(err) => return Some(Err(err)),
-        }
-        if start.elapsed() > timeout {
-            return None;
-        }
-        thread::sleep(Duration::from_millis(10));
+    match child.wait_timeout(timeout) {
+        Ok(Some(status)) => Some(Ok(status)),
+        Ok(None) => None,
+        Err(err) => Some(Err(err)),
     }
 }