diff --git a/provider_openai.go b/provider_openai.go
index 9103be1..106126f 100644
--- a/provider_openai.go
+++ b/provider_openai.go
@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"strings"
 )
 
 // openaiAdapter is the providerAdapter for provider type "openai": OpenAI
@@ -101,7 +102,16 @@ func (a *openaiAdapter) chatCompletion(ctx context.Context, w http.ResponseWrite
 		}
 	}
 
-	resp, err := upstreamJSON(ctx, a.client, http.MethodPost, a.baseURL+"/v1/chat/completions", a.requestHeaders(true), req)
+	hdr := a.requestHeaders(true)
+	if streaming {
+		// Ask the upstream for an SSE response. The Content-Type check below
+		// still decides which forwarding path runs — an upstream that
+		// ignores stream=true may answer with JSON regardless of this
+		// header.
+		hdr.Set("Accept", "text/event-stream")
+	}
+
+	resp, err := upstreamJSON(ctx, a.client, http.MethodPost, a.baseURL+"/v1/chat/completions", hdr, req)
 	if err != nil {
 		return usage{}, err
 	}
@@ -111,9 +121,13 @@ func (a *openaiAdapter) chatCompletion(ctx context.Context, w http.ResponseWrite
 		return usage{}, newProviderHTTPError(resp)
 	}
 
-	if streaming {
+	if streaming && strings.Contains(resp.Header.Get("Content-Type"), "text/event-stream") {
 		return a.forwardStream(w, resp.Body, clientAskedUsage)
 	}
+	// Either a non-streaming request, or a streaming one whose upstream
+	// ignored stream=true and answered with a normal JSON body: forward it
+	// through the same verbatim passthrough non-streaming responses use,
+	// instead of emitting an empty SSE stream for a body that was never SSE.
 	return a.forwardJSON(w, resp)
 }
 
@@ -151,7 +165,7 @@ type chatUsagePayload struct {
 func (a *openaiAdapter) forwardJSON(w http.ResponseWriter, resp *http.Response) (usage, error) {
 	body, err := io.ReadAll(io.LimitReader(resp.Body, maxResponseBytes))
 	if err != nil {
-		return usage{}, fmt.Errorf("%w: read response body: %v", errUpstream, err) //nolint:errorlint
+		return usage{}, fmt.Errorf("%w: read response body: %w", errUpstream, err)
 	}
 
 	var payload chatUsagePayload
@@ -176,7 +190,11 @@ const sseDoneMarker = "[DONE]"
 // the returned usage either way, but only forwarded to the client when
 // clientAskedUsage is true — otherwise it is dropped, since the client
 // never asked for it and chatCompletion only requested it upstream for
-// accounting. The "[DONE]" sentinel is always forwarded.
+// accounting. The "[DONE]" sentinel is always forwarded. On a mid-stream
+// read error (upstream drops the connection, client context is canceled),
+// forwardStream still returns whatever usage it had already captured
+// alongside the error — a usage chunk that arrived before the break must
+// not be discarded and billed as zero.
 func (a *openaiAdapter) forwardStream(w http.ResponseWriter, body io.Reader, clientAskedUsage bool) (usage, error) {
 	sw := newSSEWriter(w)
 	var u usage
@@ -201,7 +219,10 @@ func (a *openaiAdapter) forwardStream(w http.ResponseWriter, body io.Reader, cli
 		return sw.writeData(ev.data)
 	})
 	if err != nil {
-		return usage{}, fmt.Errorf("%w: stream: %v", errUpstream, err) //nolint:errorlint
+		// u carries whatever usage was already captured before the stream
+		// broke — a usage chunk that arrived, then a dropped connection,
+		// must not report as zero and get treated as an unbilled request.
+		return u, fmt.Errorf("%w: stream: %w", errUpstream, err)
 	}
 	return u, nil
 }
@@ -228,12 +249,12 @@ func (a *openaiAdapter) listModels(ctx context.Context) ([]string, error) {
 
 	body, err := io.ReadAll(io.LimitReader(resp.Body, maxResponseBytes))
 	if err != nil {
-		return nil, fmt.Errorf("%w: read response body: %v", errUpstream, err) //nolint:errorlint
+		return nil, fmt.Errorf("%w: read response body: %w", errUpstream, err)
 	}
 
 	var payload modelsPayload
 	if err := json.Unmarshal(body, &payload); err != nil {
-		return nil, fmt.Errorf("%w: decode models response: %v", errUpstream, err) //nolint:errorlint
+		return nil, fmt.Errorf("%w: decode models response: %w", errUpstream, err)
 	}
 
 	ids := make([]string, len(payload.Data))
diff --git a/provider_openai_test.go b/provider_openai_test.go
index d371cf6..a48ee22 100644
--- a/provider_openai_test.go
+++ b/provider_openai_test.go
@@ -247,6 +247,123 @@ func TestOpenAIAdapter_ChatCompletion_Streaming_ClientAskedUsage(t *testing.T) {
 	}
 }
 
+// TestOpenAIAdapter_ChatCompletion_Streaming_UsageCapturedBeforeAbruptClose
+// proves that when the upstream sends the usage-only chunk and then the
+// connection drops (no [DONE], no clean EOF), chatCompletion still returns
+// the usage it had already captured alongside the error — a dropped
+// connection after usage arrived must not read as zero usage.
+func TestOpenAIAdapter_ChatCompletion_Streaming_UsageCapturedBeforeAbruptClose(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		w.WriteHeader(http.StatusOK)
+		fmt.Fprintf(w, "data: %s\n\n", streamFrames[3]) // usage-only chunk: {prompt:7 completion:9}
+		w.(http.Flusher).Flush()
+
+		hj, ok := w.(http.Hijacker)
+		if !ok {
+			t.Fatalf("httptest ResponseWriter does not implement http.Hijacker")
+		}
+		conn, _, err := hj.Hijack()
+		if err != nil {
+			t.Fatalf("hijack: %v", err)
+		}
+		// Abrupt close: no [DONE], no terminating chunk — the client must
+		// see a genuine read error, not a clean EOF.
+		_ = conn.Close()
+	}))
+	defer srv.Close()
+
+	a := newOpenAIAdapter("p1", srv.URL, "sk-test")
+	rec := newRecordingResponseWriter()
+
+	req := map[string]any{"model": "gpt-5", "stream": true}
+	u, err := a.chatCompletion(context.Background(), rec, req)
+	if err == nil {
+		t.Fatalf("chatCompletion: want error from abrupt close, got nil (usage=%+v)", u)
+	}
+	if !errors.Is(err, errUpstream) {
+		t.Errorf("err = %v, want errors.Is(err, errUpstream) == true", err)
+	}
+	if u.prompt != 7 || u.completion != 9 {
+		t.Errorf("usage = %+v, want {prompt:7 completion:9} (captured before the drop)", u)
+	}
+}
+
+// TestOpenAIAdapter_ChatCompletion_ContextCanceledMidStream proves that
+// canceling the caller's context while a streaming response is still being
+// read surfaces as errors.Is(err, context.Canceled) at chatCompletion's
+// caller — required so Task 12 can tell a client hang-up apart from a real
+// upstream failure.
+func TestOpenAIAdapter_ChatCompletion_ContextCanceledMidStream(t *testing.T) {
+	block := make(chan struct{})
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		w.WriteHeader(http.StatusOK)
+		fmt.Fprintf(w, "data: %s\n\n", streamFrames[0])
+		w.(http.Flusher).Flush()
+		<-block // hold the connection open past the context cancellation below
+	}))
+	defer srv.Close()
+	defer close(block) // unblock the handler before srv.Close() waits on it
+
+	a := newOpenAIAdapter("p1", srv.URL, "sk-test")
+	rec := newRecordingResponseWriter()
+
+	ctx, cancel := context.WithCancel(context.Background())
+	go func() {
+		time.Sleep(50 * time.Millisecond)
+		cancel()
+	}()
+
+	_, err := a.chatCompletion(ctx, rec, map[string]any{"model": "gpt-5", "stream": true})
+	if err == nil {
+		t.Fatalf("chatCompletion: want error from context cancellation, got nil")
+	}
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("err = %v, want errors.Is(err, context.Canceled) == true", err)
+	}
+}
+
+// TestOpenAIAdapter_ChatCompletion_Streaming_UpstreamIgnoresStreamFallsBackToJSON
+// proves that when a streaming request's upstream answers with a plain
+// JSON body instead of SSE (Content-Type not text/event-stream), and sends
+// "Accept: text/event-stream" on the outgoing request, chatCompletion falls
+// back to the non-streaming passthrough — the client gets the JSON body
+// verbatim and usage is still extracted — instead of emitting an empty or
+// broken SSE stream.
+func TestOpenAIAdapter_ChatCompletion_Streaming_UpstreamIgnoresStreamFallsBackToJSON(t *testing.T) {
+	const respBody = `{"id":"chatcmpl-1","choices":[{"message":{"role":"assistant","content":"hi"}}],"usage":{"prompt_tokens":3,"completion_tokens":4}}`
+
+	var gotAccept string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotAccept = r.Header.Get("Accept")
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(respBody))
+	}))
+	defer srv.Close()
+
+	a := newOpenAIAdapter("p1", srv.URL, "sk-test")
+	rec := httptest.NewRecorder()
+
+	u, err := a.chatCompletion(context.Background(), rec, map[string]any{"model": "gpt-5", "stream": true})
+	if err != nil {
+		t.Fatalf("chatCompletion: %v", err)
+	}
+	if u.prompt != 3 || u.completion != 4 {
+		t.Errorf("usage = %+v, want {prompt:3 completion:4}", u)
+	}
+	if rec.Body.String() != respBody {
+		t.Errorf("body = %q, want %q (JSON fallback, verbatim passthrough)", rec.Body.String(), respBody)
+	}
+	if got := rec.Header().Get("Content-Type"); got != "application/json" {
+		t.Errorf("Content-Type = %q, want application/json", got)
+	}
+	if gotAccept != "text/event-stream" {
+		t.Errorf("Accept header sent = %q, want %q", gotAccept, "text/event-stream")
+	}
+}
+
 // TestOpenAIAdapter_ChatCompletion_Upstream429 proves a non-2xx upstream
 // response returns a *providerHTTPError carrying the status and body,
 // with nothing written to the client — the caller decides envelope vs
diff --git a/providers.go b/providers.go
index dbd354c..207528f 100644
--- a/providers.go
+++ b/providers.go
@@ -115,7 +115,17 @@ type providerAdapter interface {
 // pool, since every request from one adapter targets the same upstream
 // host.
 func newAdapterHTTPClient() *http.Client {
-	tr := http.DefaultTransport.(*http.Transport).Clone() //nolint:forcetypeassert // http.DefaultTransport is always *http.Transport
+	tr, ok := http.DefaultTransport.(*http.Transport)
+	if ok {
+		tr = tr.Clone()
+	} else {
+		// Defensive fallback: something in the shared Traefik process
+		// replaced http.DefaultTransport with a type that is not
+		// *http.Transport. A fresh zero-value Transport (Go's own defaults)
+		// is safer than panicking at adapter construction over a type
+		// assertion this package does not control.
+		tr = &http.Transport{}
+	}
 	tr.MaxIdleConnsPerHost = adapterIdleConnsPerHost
 	return &http.Client{Transport: tr}
 }
@@ -135,14 +145,14 @@ func upstreamJSON(ctx context.Context, client *http.Client, method, url string,
 	if body != nil {
 		b, err := json.Marshal(body)
 		if err != nil {
-			return nil, fmt.Errorf("%w: encode request body: %v", errUpstream, err) //nolint:errorlint // errUpstream is a sentinel wrapped for errors.Is, %v is fine for the inner cause
+			return nil, fmt.Errorf("%w: encode request body: %w", errUpstream, err)
 		}
 		r = bytes.NewReader(b)
 	}
 
 	req, err := http.NewRequestWithContext(ctx, method, url, r)
 	if err != nil {
-		return nil, fmt.Errorf("%w: build request: %v", errUpstream, err) //nolint:errorlint
+		return nil, fmt.Errorf("%w: build request: %w", errUpstream, err)
 	}
 	for k, vs := range hdr {
 		for _, v := range vs {
@@ -155,7 +165,7 @@ func upstreamJSON(ctx context.Context, client *http.Client, method, url string,
 
 	resp, err := client.Do(req) //nolint:bodyclose // caller closes resp.Body; upstreamJSON hands the response, not its lifecycle, back
 	if err != nil {
-		return nil, fmt.Errorf("%w: %v", errUpstream, err) //nolint:errorlint
+		return nil, fmt.Errorf("%w: %w", errUpstream, err)
 	}
 	return resp, nil
 }