diff --git a/src/harness/structured/mod.rs b/src/harness/structured/mod.rs
index 8eec80c..525cb30 100644
--- a/src/harness/structured/mod.rs
+++ b/src/harness/structured/mod.rs
@@ -184,6 +184,34 @@ impl StructuredExtractor {
 
     fn extract_provider_schema(&self, response: &ModelResponse) -> Result<StructuredOutput> {
         let raw = response.text();
+
+        // A reasoning model can spend its entire output budget thinking and
+        // return no content at all. The provider says so plainly through
+        // `finish_reason: "length"`, but a naive parse of the empty string
+        // reports "expected value at line 1 column 1" — which sends the reader
+        // hunting for a malformed response that was never sent, and hides the
+        // one-line fix of raising `max_tokens` or capping reasoning.
+        if raw.trim().is_empty() {
+            let truncated = response
+                .finish_reason
+                .as_deref()
+                .is_some_and(|reason| reason == "length");
+            return Err(TinyAgentsError::StructuredOutput(if truncated {
+                format!(
+                    "schema '{}': the model returned no content because it hit its output limit \
+                     (finish_reason = \"length\"). Reasoning models can consume the whole budget \
+                     before emitting an answer: raise `max_tokens`, or cap/disable reasoning.",
+                    self.schema_name
+                )
+            } else {
+                format!(
+                    "schema '{}': the model returned no content (finish_reason = {:?})",
+                    self.schema_name,
+                    response.finish_reason.as_deref().unwrap_or("unknown")
+                )
+            }));
+        }
+
         let value: Value = serde_json::from_str(&raw).map_err(|e| {
             TinyAgentsError::StructuredOutput(format!(
                 "schema '{}': response text is not valid JSON: {e}",
diff --git a/src/harness/structured/test.rs b/src/harness/structured/test.rs
index cf0b0c6..0145d2b 100644
--- a/src/harness/structured/test.rs
+++ b/src/harness/structured/test.rs
@@ -111,3 +111,52 @@ fn structured_output_parse_deserialises() {
     let parsed: Answer = output.parse().unwrap();
     assert_eq!(parsed.value, "hello");
 }
+
+// -- truncation diagnostics --
+
+/// A response whose content is `text`, stopped for `finish_reason`.
+fn stopped(text: &str, finish_reason: &str) -> ModelResponse {
+    ModelResponse::assistant(text).with_finish_reason(finish_reason)
+}
+
+#[test]
+fn empty_content_stopped_for_length_reports_the_output_limit() {
+    // A reasoning model can spend its whole output budget thinking and return
+    // no content. Parsing the empty string reports "expected value at line 1
+    // column 1", which sends the reader hunting for a malformed response that
+    // was never sent, and hides the one-line fix.
+    let extractor =
+        StructuredExtractor::new(StructuredStrategy::ProviderSchema, "review", json!({}));
+    let err = extractor
+        .extract(&stopped("", "length"))
+        .expect_err("empty content is not extractable");
+
+    let message = err.to_string();
+    assert!(message.contains("output limit"), "{message}");
+    assert!(message.contains("max_tokens"), "{message}");
+    assert!(
+        !message.contains("line 1 column 1"),
+        "the misleading parse error must not survive: {message}"
+    );
+}
+
+#[test]
+fn empty_content_stopped_normally_reports_what_is_known() {
+    let extractor =
+        StructuredExtractor::new(StructuredStrategy::ProviderSchema, "review", json!({}));
+    let err = extractor
+        .extract(&stopped("   ", "stop"))
+        .expect_err("empty content is not extractable");
+    assert!(err.to_string().contains("no content"), "{err}");
+}
+
+#[test]
+fn genuinely_malformed_json_still_reports_a_parse_error() {
+    // The new branch must not swallow the case it was not written for.
+    let extractor =
+        StructuredExtractor::new(StructuredStrategy::ProviderSchema, "review", json!({}));
+    let err = extractor
+        .extract(&stopped("not json at all", "stop"))
+        .expect_err("malformed content is not extractable");
+    assert!(err.to_string().contains("not valid JSON"), "{err}");
+}