Ollama: only treat a 404 as model-not-found when the body says so

A 404 from a proxy without /api/chat, or any other 404 error, was
reported as the configured model being missing. Generate now checks the
{"error": ...} payload for Ollama's model-not-found message naming the
configured model (double- or single-quoted); any other 404 stays a
ProviderError with its redacted body excerpt.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

internal/ai/llm/ollama/ollama.go

@@ -138,9 +138,9 @@ func (p *Provider) Generate(ctx context.Context, req llm.Request) (llm.Response,
 	raw, err := p.do(ctx, http.MethodPost, "/api/chat", payload)
 	if err != nil {
 		var pe *llm.ProviderError
-		if errors.As(err, &pe) && pe.Status() == http.StatusNotFound {
+		if errors.As(err, &pe) && pe.Status() == http.StatusNotFound && p.modelNotFound(pe.Body()) {
-			// Ollama answers 404 when the model is not pulled. The body is
+			// The body is not quoted: the model name is all the caller
-			// not quoted: the model name is all the caller needs.
+			// needs. Any other 404 keeps its body excerpt below.
 			return llm.Response{}, fmt.Errorf("ollama: generate: model %q not found: %w", p.model, llm.NewProviderError(pe.Status(), nil, ""))
 		}
 		return llm.Response{}, fmt.Errorf("ollama: generate: %w", err)
@@ -169,6 +169,21 @@ func (p *Provider) Generate(ctx context.Context, req llm.Request) (llm.Response,
 	return resp, fmt.Errorf("ollama: generate: reply is not a JSON object: %w", llm.ErrBadOutput)
 }
 
+// modelNotFound reports whether body is Ollama's error for the configured
+// model not being pulled: {"error": "model \"name\" not found, ..."}, or
+// with the name in single quotes as newer servers send it. A 404 from
+// anything else, such as a proxy without /api/chat, is not.
+func (p *Provider) modelNotFound(body string) bool {
+	var e struct {
+		Error string `json:"error"`
+	}
+	if err := json.Unmarshal([]byte(body), &e); err != nil {
+		return false
+	}
+	return strings.HasPrefix(e.Error, fmt.Sprintf("model %q not found", p.model)) ||
+		strings.HasPrefix(e.Error, fmt.Sprintf("model '%s' not found", p.model))
+}
+
 // isJSONObject reports whether s is exactly one JSON object.
 func isJSONObject(s string) bool {
 	return strings.HasPrefix(s, "{") && json.Valid([]byte(s))

internal/ai/llm/ollama/ollama_test.go

@@ -327,6 +327,49 @@ func TestGenerateModelMissing(t *testing.T) {
 	}
 }
 
+func TestGenerateNotFoundUnrelatedToModel(t *testing.T) {
+	for name, body := range map[string]string{
+		"proxy page":  "404 page not found",
+		"other error": `{"error":"no route for /api/chat"}`,
+		"other model": `{"error":"model \"llama3\" not found, try pulling it first"}`,
+		"empty":       "",
+	} {
+		t.Run(name, func(t *testing.T) {
+			s := newOllamaServer(t, reply{status: 404, body: body})
+			p, err := New(Config{BaseURL: s.URL, Model: "llama9"})
+			if err != nil {
+				t.Fatal(err)
+			}
+			_, err = p.Generate(context.Background(), textRequest())
+			var pe *llm.ProviderError
+			if !errors.Is(err, llm.ErrProvider) || !errors.As(err, &pe) || pe.Status() != http.StatusNotFound {
+				t.Fatalf("err = %v, want a ProviderError for HTTP 404", err)
+			}
+			if strings.Contains(err.Error(), "not found") && !strings.Contains(body, "not found") {
+				t.Fatalf("Error() = %q claims something was not found", err)
+			}
+			if strings.Contains(err.Error(), "llama9") {
+				t.Fatalf("Error() = %q claims the model was not found", err)
+			}
+			if pe.Body() != strings.TrimSpace(body) {
+				t.Fatalf("Body() = %q, want the response body %q", pe.Body(), body)
+			}
+		})
+	}
+}
+
+func TestGenerateModelMissingSingleQuoted(t *testing.T) {
+	s := newOllamaServer(t, reply{status: 404, body: `{"error":"model 'llama9' not found"}`})
+	p, err := New(Config{BaseURL: s.URL, Model: "llama9"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = p.Generate(context.Background(), textRequest())
+	if want := `ollama: generate: model "llama9" not found: llm: provider error: HTTP 404`; err == nil || err.Error() != want {
+		t.Fatalf("err = %v, want %q", err, want)
+	}
+}
+
 func TestListModels(t *testing.T) {
 	s := newOllamaServer(t, reply{status: 200, fixture: "tags.json"})
 	models, err := newProvider(t, s.URL).ListModels(context.Background())