diff --git a/src/harness/model/mod.rs b/src/harness/model/mod.rs
index aa24cb1..80bfd24 100644
--- a/src/harness/model/mod.rs
+++ b/src/harness/model/mod.rs
@@ -694,8 +694,13 @@ pub struct StreamAccumulator {
     usage: Option<Usage>,
     /// Authoritative final response, when a `Completed` item was seen.
     completed: Option<ModelResponse>,
-    /// Terminal error message, when a failure item was seen.
+    /// Terminal error message, when an unstructured `Failed` item was seen.
     failed: Option<String>,
+    /// Terminal structured provider failure, when a `ProviderFailed` item was
+    /// seen. Kept as the full struct (not stringified) so `finish()` can return
+    /// [`crate::error::TinyAgentsError::Provider`] and preserve the
+    /// status/code/`retryable` classification the retry layer needs.
+    failed_provider: Option<ProviderError>,
 }
 
 impl StreamAccumulator {
@@ -732,16 +737,11 @@ impl StreamAccumulator {
                 self.failed = Some(message.clone());
             }
             ModelStreamItem::ProviderFailed(error) => {
-                self.failed = Some(format!(
-                    "{} provider error{}: {}",
-                    error.provider,
-                    error
-                        .code
-                        .as_deref()
-                        .map(|code| format!(" ({code})"))
-                        .unwrap_or_default(),
-                    error.message
-                ));
+                // Retain the structured error rather than stringifying it, so
+                // `finish()` can surface `TinyAgentsError::Provider` and the retry
+                // layer sees the real status/code/`retryable` (a permanent 401 /
+                // `insufficient_quota` / 400 must not be retried as transient).
+                self.failed_provider = Some(error.clone());
             }
         }
     }
@@ -753,10 +753,10 @@ impl StreamAccumulator {
     fn push_tool_chunk(&mut self, call_id: &str, content: &str, tool_name: Option<&str>) {
         if let Some(entry) = self.tool_chunks.iter_mut().find(|(id, ..)| id == call_id) {
             entry.1.push_str(content);
-            if entry.2.is_none() {
-                if let Some(name) = tool_name.filter(|n| !n.is_empty()) {
-                    entry.2 = Some(name.to_string());
-                }
+            if entry.2.is_none()
+                && let Some(name) = tool_name.filter(|n| !n.is_empty())
+            {
+                entry.2 = Some(name.to_string());
             }
         } else {
             self.tool_chunks.push((
@@ -767,10 +767,10 @@ impl StreamAccumulator {
         }
     }
 
-    /// Returns `true` when a terminal item (`Completed` or `Failed`) has been
-    /// folded in.
+    /// Returns `true` when a terminal item (`Completed`, `Failed`, or
+    /// `ProviderFailed`) has been folded in.
     pub fn is_terminal(&self) -> bool {
-        self.completed.is_some() || self.failed.is_some()
+        self.completed.is_some() || self.failed.is_some() || self.failed_provider.is_some()
     }
 
     /// Returns the accumulated reasoning/thinking text streamed so far (the
@@ -783,9 +783,16 @@ impl StreamAccumulator {
     ///
     /// # Errors
     ///
-    /// Returns [`crate::error::TinyAgentsError::Model`] when a
-    /// [`ModelStreamItem::Failed`] item was folded in.
+    /// Returns [`crate::error::TinyAgentsError::Provider`] when a
+    /// [`ModelStreamItem::ProviderFailed`] item was folded in — preserving the
+    /// structured status/code/`retryable` so a permanent failure is not retried
+    /// as transient — or [`crate::error::TinyAgentsError::Model`] when an
+    /// unstructured [`ModelStreamItem::Failed`] item was folded in.
     pub fn finish(self) -> Result<ModelResponse> {
+        if let Some(error) = self.failed_provider {
+            return Err(crate::error::TinyAgentsError::Provider(Box::new(error)));
+        }
+
         if let Some(message) = self.failed {
             return Err(crate::error::TinyAgentsError::Model(message));
         }
diff --git a/src/harness/model/test.rs b/src/harness/model/test.rs
index 64a86c5..63e5036 100644
--- a/src/harness/model/test.rs
+++ b/src/harness/model/test.rs
@@ -684,6 +684,59 @@ fn finish_backfills_usage_from_stream_delta_when_completed_lacks_it() {
     assert_eq!(finished.message.usage, Some(Usage::new(4, 6)));
 }
 
+#[test]
+fn finish_preserves_provider_error_classification_from_provider_failed() {
+    // A streamed `ProviderFailed` carrying a non-retryable provider error (401
+    // auth) must surface as `TinyAgentsError::Provider` with the struct intact —
+    // not stringified into `Model` — so `is_retryable` classifies it as permanent
+    // instead of retrying + fallback-churning it as transient.
+    use crate::error::TinyAgentsError;
+    use crate::harness::retry::is_retryable;
+
+    let mut acc = StreamAccumulator::new();
+    acc.push(&ModelStreamItem::Started);
+    acc.push(&ModelStreamItem::ProviderFailed(ProviderError {
+        provider: "openai".into(),
+        status: Some(401),
+        code: Some("invalid_api_key".into()),
+        message: "Incorrect API key provided".into(),
+        retryable: false,
+        ..ProviderError::default()
+    }));
+
+    assert!(acc.is_terminal());
+    let err = acc.finish().unwrap_err();
+    match &err {
+        TinyAgentsError::Provider(error) => {
+            assert_eq!(error.status, Some(401));
+            assert_eq!(error.code.as_deref(), Some("invalid_api_key"));
+            assert!(!error.retryable);
+        }
+        other => panic!("expected Provider error, got {other:?}"),
+    }
+    assert!(
+        !is_retryable(&err),
+        "a permanent streamed provider failure must not be retried as transient"
+    );
+}
+
+#[test]
+fn finish_maps_unstructured_failed_to_model_error() {
+    // The unstructured `Failed(String)` path stays `TinyAgentsError::Model` — no
+    // structured detail to classify from, so the retry layer treats it as a
+    // transient transport/parse failure.
+    use crate::error::TinyAgentsError;
+
+    let mut acc = StreamAccumulator::new();
+    acc.push(&ModelStreamItem::Failed("stream broke".into()));
+
+    assert!(acc.is_terminal());
+    match acc.finish().unwrap_err() {
+        TinyAgentsError::Model(message) => assert_eq!(message, "stream broke"),
+        other => panic!("expected Model error, got {other:?}"),
+    }
+}
+
 #[test]
 fn finish_names_reconstructed_tool_call_from_the_call_opening_delta_name() {
     // A call-opening delta carries the tool name (no args yet); subsequent
diff --git a/src/harness/providers/openai/test.rs b/src/harness/providers/openai/test.rs
index 0868568..3ca9066 100644
--- a/src/harness/providers/openai/test.rs
+++ b/src/harness/providers/openai/test.rs
@@ -513,7 +513,11 @@ fn provider_spec_builds_compatible_model() {
 }
 
 #[test]
-fn provider_failed_stream_item_finishes_as_model_error() {
+fn provider_failed_stream_item_finishes_as_provider_error() {
+    // A streamed `ProviderFailed` must finish as `TinyAgentsError::Provider` with
+    // the structured error preserved (status/code/`retryable`), so the retry
+    // layer classifies it from `retryable` instead of the old behavior that
+    // stringified it into `Model` and always retried it as transient.
     let mut accumulator = StreamAccumulator::new();
     accumulator.push(&ModelStreamItem::ProviderFailed(ProviderError {
         provider: "groq".to_string(),
@@ -525,8 +529,16 @@ fn provider_failed_stream_item_finishes_as_model_error() {
         raw: None,
     }));
 
-    let error = accumulator.finish().unwrap_err().to_string();
-    assert!(error.contains("groq provider error (rate_limit): too many requests"));
+    match accumulator.finish().unwrap_err() {
+        TinyAgentsError::Provider(error) => {
+            assert_eq!(error.provider, "groq");
+            assert_eq!(error.status, Some(429));
+            assert_eq!(error.code.as_deref(), Some("rate_limit"));
+            assert_eq!(error.message, "too many requests");
+            assert!(error.retryable);
+        }
+        other => panic!("expected Provider error, got {other:?}"),
+    }
 }
 
 #[tokio::test]
diff --git a/src/harness/stream/mod.rs b/src/harness/stream/mod.rs
index 14411f0..1712142 100644
--- a/src/harness/stream/mod.rs
+++ b/src/harness/stream/mod.rs
@@ -143,6 +143,5 @@ pub fn stream(chunks: &[StreamChunk], modes: &[StreamMode]) -> Vec<StreamChunk>
         .collect()
 }
 
-#[cfg(test)]
 #[cfg(test)]
 mod test;
diff --git a/tests/e2e_harness_provider_contracts.rs b/tests/e2e_harness_provider_contracts.rs
index a4b09b5..5f3baee 100644
--- a/tests/e2e_harness_provider_contracts.rs
+++ b/tests/e2e_harness_provider_contracts.rs
@@ -281,9 +281,14 @@ async fn model_request_response_registry_and_stream_contracts_are_stable() {
         },
     )]));
     let err = collect_model_stream(failed_stream).await.unwrap_err();
+    // A streamed `ProviderFailed` now surfaces as a structured
+    // `TinyAgentsError::Provider` whose `Display` renders the provider, HTTP
+    // status, code, and message (preserving the classification the retry layer
+    // needs) rather than the old flattened "<provider> provider error" string.
     assert!(
         err.to_string()
-            .contains("mock provider error (internal): nope")
+            .contains("mock returned HTTP 500 (internal): nope"),
+        "unexpected error string: {err}"
     );
 }