diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 7012da2b..337a077f 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -116,6 +116,7 @@ jobs:
               - '**/*.sh'
               - '**/*.ps1'
               - '**/*.feature'
+              - 'tests/e2e-cucumber/expectations.toml'
               - 'install*'
               - '.github/workflows/**'
             # THIRD_PARTY_NOTICES.txt staleness gate: anything that changes the
diff --git a/tests/e2e-cucumber/src/lib.rs b/tests/e2e-cucumber/src/lib.rs
index 07471e93..3fcdcec2 100644
--- a/tests/e2e-cucumber/src/lib.rs
+++ b/tests/e2e-cucumber/src/lib.rs
@@ -6,8 +6,37 @@ pub mod capability;
 pub mod expectation;
 pub mod mock_server;
 
+pub fn chat_response_is_successful(response: &serde_json::Value) -> bool {
+    response
+        .get("choices")
+        .and_then(serde_json::Value::as_array)
+        .is_some_and(|choices| !choices.is_empty())
+}
+
 // The report generator lives in its own lean crate (only maud + serde) so xtask
 // can reuse it without pulling this harness's heavy tree. Re-export it under the
 // original path so `e2e_cucumber::report::{generate, evaluate_xfail}` call sites
 // keep working.
 pub use e2e_report as report;
+
+#[cfg(test)]
+mod tests {
+    use super::chat_response_is_successful;
+
+    #[test]
+    fn chat_success_requires_non_empty_choices_array() {
+        assert!(!chat_response_is_successful(&serde_json::json!({})));
+        assert!(!chat_response_is_successful(
+            &serde_json::json!({"choices": null})
+        ));
+        assert!(!chat_response_is_successful(
+            &serde_json::json!({"choices": {}})
+        ));
+        assert!(!chat_response_is_successful(
+            &serde_json::json!({"choices": []})
+        ));
+        assert!(chat_response_is_successful(
+            &serde_json::json!({"choices": [{}]})
+        ));
+    }
+}
diff --git a/tests/e2e-cucumber/tests/e2e/chat_steps.rs b/tests/e2e-cucumber/tests/e2e/chat_steps.rs
index d2c08cdd..b781de12 100644
--- a/tests/e2e-cucumber/tests/e2e/chat_steps.rs
+++ b/tests/e2e-cucumber/tests/e2e/chat_steps.rs
@@ -41,8 +41,18 @@ async fn user_offered_endpoint(_world: &mut E2eWorld) {}
 #[when("a chat request with tool definitions is sent")]
 async fn send_chat_with_tools(world: &mut E2eWorld) {
     let endpoint = world.endpoint.as_ref().expect("no endpoint configured");
+
+    let client = reqwest::Client::builder()
+        .timeout(std::time::Duration::from_secs(
+            crate::inference_timeout_for(world),
+        ))
+        .build()
+        .expect("failed to build HTTP client");
+
     let models_url = format!("{endpoint}/models");
-    let resp: serde_json::Value = reqwest::get(&models_url)
+    let resp: serde_json::Value = client
+        .get(&models_url)
+        .send()
         .await
         .unwrap_or_else(|e| panic!("GET {models_url} failed: {e}"))
         .json()
@@ -54,7 +64,6 @@ async fn send_chat_with_tools(world: &mut E2eWorld) {
         .to_string();
 
     let chat_url = format!("{endpoint}/chat/completions");
-    let client = reqwest::Client::new();
     let chat_resp: serde_json::Value = client
         .post(&chat_url)
         .json(&serde_json::json!({
@@ -152,8 +161,8 @@ async fn assert_privacy_notice_accurate(_world: &mut E2eWorld) {
 async fn assert_chat_successful(world: &mut E2eWorld) {
     let resp = world.chat_response.as_ref().expect("no chat response");
     assert!(
-        resp.get("choices").is_some(),
-        "no choices in response: {resp}"
+        e2e_cucumber::chat_response_is_successful(resp),
+        "no non-empty choices array in response: {resp}"
     );
 }