fix(openai): keep transport failures structured instead of flattening them

src/harness/providers/openai/README.md

@@ -224,9 +224,18 @@ into a structured `ProviderError` (HTTP status, provider error code, and a
 everything else — including 401/400 — is not) and surfaced as
 `TinyAgentsError::Provider`, so `harness::retry::is_retryable` can classify
 retryability instead of retrying every provider failure indiscriminately.
-Transport-level failures (connection errors, body-read failures) have no such
-structure to preserve and surface as a plain `TinyAgentsError::Model` string
-via `provider_failure_message`. Malformed JSON bodies surface as
+A **send** failure — no final HTTP status was available, so there is nothing to
+report as one — still carries a provider, a model, and a computed `retryable`,
+so `send_checked` raises it as `TinyAgentsError::Provider` with `status: None`
+rather than flattening it; a host that classifies on the variant would
+otherwise see a connection reset as an unstructured error. `Display` is
+identical either way. Note that "no final status" is not the same as "never
+reached a server": the client keeps reqwest's default redirect policy, so a
+redirect loop or an exceeded redirect limit lands here too, after servers have
+answered with 3xx.
+
+**Body-read** failures after a 2xx (`list_models`, `invoke_responses`) remain
+plain `TinyAgentsError::Model` strings. Malformed JSON bodies surface as
 `TinyAgentsError::Serialization`.
 
 ## Operational constraints

src/harness/providers/openai/test.rs

@@ -2915,3 +2915,63 @@ fn stream_cleanup_scrubs_leaked_markup_from_live_deltas() {
     assert_eq!(completed.tool_calls().len(), 1);
     assert_eq!(completed.tool_calls()[0].name, "lookup");
 }
+
+// ── transport-level failures keep their structure ───────────────────────────
+
+/// A transport failure must reach the caller as `TinyAgentsError::Provider`.
+///
+/// It carries no HTTP status — nothing answered — but it does carry the
+/// provider, the model, and a computed `retryable`, and a host that classifies
+/// on the variant needs all three. Flattening it into
+/// `TinyAgentsError::Model(String)` sends a connection reset down the generic
+/// arm of a host's logger, where it is recorded with no status, no provider and
+/// no retryability, and the operator reading it cannot tell a dead endpoint
+/// from a rejected request.
+#[tokio::test]
+async fn a_transport_failure_is_reported_as_a_structured_provider_error() {
+    // Bind, read the port, then drop the listener: the port is now almost
+    // certainly closed, so the connect fails without waiting on a timeout.
+    let port = {
+        let listener = std::net::TcpListener::bind("127.0.0.1:0").expect("bind an ephemeral port");
+        listener.local_addr().expect("read the bound port").port()
+    };
+
+    let model = OpenAiModel::compatible_provider(
+        "OpenHuman",
+        "test-key",
+        format!("http://127.0.0.1:{port}/openai/v1"),
+        "chat-v1",
+    );
+
+    let error = ChatModel::<()>::invoke(&model, &(), ModelRequest::new(vec![Message::user("hi")]))
+        .await
+        .expect_err("a closed port cannot answer");
+
+    let TinyAgentsError::Provider(provider_error) = &error else {
+        panic!("expected a structured Provider error, got: {error:?}");
+    };
+    assert_eq!(provider_error.provider, "OpenHuman");
+    assert_eq!(provider_error.model.as_deref(), Some("chat-v1"));
+    assert!(
+        provider_error.status.is_none(),
+        "nothing answered, so there is no status to report"
+    );
+    assert!(
+        provider_error.retryable,
+        "a transport failure is worth retrying; the flag is what says so"
+    );
+
+    // The rendered text is the no-regression half: `Display for ProviderError`
+    // reproduces what `provider_failure_message` used to build by hand, and both
+    // error variants render as `model error: {0}`, so nothing a user or a log
+    // reader sees changes.
+    let rendered = error.to_string();
+    assert!(
+        rendered.starts_with("model error: OpenHuman returned: "),
+        "the user-facing wording must not change, got: {rendered}"
+    );
+    assert!(
+        rendered.contains("/openai/v1/chat/completions"),
+        "the failing URL is what makes this diagnosable, got: {rendered}"
+    );
+}

src/harness/providers/openai/transport.rs

@@ -1993,9 +1993,19 @@ impl OpenAiModel {
         url: &str,
     ) -> Result<reqwest::Response> {
         let response = builder.send().await.map_err(|e| {
-            let error =
-                self.provider_error(format!("{what} to {url} failed: {e}"), None, None, None);
-            TinyAgentsError::Model(self.provider_failure_message(&error))
+            // Structured, not flattened. A transport failure has no HTTP status,
+            // but it does have a provider, a model, and a computed `retryable` —
+            // and a host that matches on `TinyAgentsError::Provider` to log
+            // those fields would otherwise fall through to its generic arm and
+            // record a connection reset as "(non-provider error)" with nothing
+            // to act on. `Display` renders identically either way, so only the
+            // machine-readable half changes.
+            TinyAgentsError::from_provider_error(self.provider_error(
+                format!("{what} to {url} failed: {e}"),
+                None,
+                None,
+                None,
+            ))
         })?;
 
         let status = response.status();
@@ -2133,23 +2143,6 @@ impl OpenAiModel {
         }
     }
 
-    fn provider_failure_message(&self, error: &ProviderError) -> String {
-        format!(
-            "{} returned{}{}: {}",
-            error.provider,
-            error
-                .status
-                .map(|status| format!(" HTTP {status}"))
-                .unwrap_or_default(),
-            error
-                .code
-                .as_deref()
-                .map(|code| format!(" ({code})"))
-                .unwrap_or_default(),
-            error.message
-        )
-    }
-
     /// Decodes a non-2xx body into a structured [`ProviderError`].
     ///
     /// Two classifications are applied on top of the raw decode: