Add model_discovery backend option to disable /v1/models probing

README.md

@@ -44,6 +44,8 @@ backends:
     type: gcp_openai
     project: your-google-cloud-project
     location: global
+    # Vertex AI's OpenAI-compatible endpoint does not provide /v1/models.
+    model_discovery: false
     auth:
       type: google_adc
     models:

docs/EXAMPLES.md

@@ -41,6 +41,8 @@ backends:
     type: gcp_openai
     project: example
     location: global
+    # Vertex AI's OpenAI-compatible endpoint does not provide /v1/models.
+    model_discovery: false
     auth:
       type: google_adc
     models:
@@ -202,6 +204,8 @@ backends:
     type: gcp_openai
     project: example
     location: global
+    # Vertex AI's OpenAI-compatible endpoint does not provide /v1/models.
+    model_discovery: false
     auth:
       type: google_adc
     models:

docs/USAGE.md

@@ -95,14 +95,14 @@ The app uses the TUI when stdout is an interactive terminal and `--headless` is
 The TUI has the following views, accessible with **Tab**, **←/→**, or **h/l**:
 
 - **Stats** – Displays per-model statistics and recent requests (the default view).
-- **Models** – Queries every upstream backend's `/models` endpoint and compares the response with the configured models.
+- **Models** – Queries each enabled upstream backend's `/models` endpoint and compares the response with the configured models.
 - **Test** – Lists all configured models. Use **↑/↓** to select a model and **Enter** to send a test request. The response is displayed inline.
 
 Test requests go through the proxy like any other request, so they count towards stats and token usage.
 
 ### Model Discovery Diagnostics
 
-Opening the **Models** tab calls the OpenAI-compatible `/models` endpoint on each configured upstream backend. The request uses the backend's configured base URL, authentication, HTTP client, and TLS settings. It does not compare against the proxy's own `GET /v1/models` endpoint, because that endpoint is generated from the config itself.
+Opening the **Models** tab calls the OpenAI-compatible `/models` endpoint on each enabled upstream backend. The request uses the backend's configured base URL, authentication, HTTP client, and TLS settings. It does not compare against the proxy's own `GET /v1/models` endpoint, because that endpoint is generated from the config itself. Set a backend's `model_discovery: false` when its provider does not support the endpoint; the TUI shows it as disabled instead of sending a request.
 
 The combined list distinguishes the source and consistency of every model:
 
@@ -130,6 +130,7 @@ Each backend declares where requests go, how they authenticate, and which models
 | `project` | yes* | Google Cloud project for `gcp_openai` when `base_url` is omitted. |
 | `location` | yes* | Google Cloud location for `gcp_openai` when `base_url` is omitted. Defaults to `global` when project is configured. |
 | `insecure_skip_verify` | no | Disables TLS certificate verification for backend API calls. Defaults to `false`. |
+| `model_discovery` | no | Whether the Models TUI tab queries this backend's upstream `/v1/models` endpoint. Defaults to `true`; set to `false` for providers that do not support it. |
 | `auth` | yes | Auth configuration. |
 | `models` | yes | `all` or a list of model entries. |
 

internal/config/config.go

@@ -45,10 +45,17 @@ type BackendConfig struct {
 	Project            string        `yaml:"project"`
 	Location           string        `yaml:"location"`
 	InsecureSkipVerify bool          `yaml:"insecure_skip_verify"`
+	ModelDiscovery     *bool         `yaml:"model_discovery"`
 	Auth               AuthConfig    `yaml:"auth"`
 	Models             BackendModels `yaml:"models"`
 }
 
+// ModelDiscoveryEnabled returns whether the Models TUI tab should query this
+// backend's upstream /v1/models endpoint. It defaults to true.
+func (bc *BackendConfig) ModelDiscoveryEnabled() bool {
+	return bc.ModelDiscovery == nil || *bc.ModelDiscovery
+}
+
 // AuthConfig describes how to authenticate against a backend.
 type AuthConfig struct {
 	Type               string   `yaml:"type"` // none, bearer, google_adc, oauth_client_credentials

internal/config/config_test.go

@@ -79,6 +79,34 @@ ui:
 	}
 }
 
+func TestLoadBackendModelDiscovery(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "config.yaml")
+	if err := os.WriteFile(path, []byte(`
+backends:
+  - name: google
+    type: gcp_openai
+    project: example
+    model_discovery: false
+    auth:
+      type: google_adc
+    models:
+      - id: gemini
+`), 0600); err != nil {
+		t.Fatal(err)
+	}
+
+	cfg, err := Load(Flags{ConfigPath: path})
+	if err != nil {
+		t.Fatalf("Load returned error: %v", err)
+	}
+	if cfg.Backends[0].ModelDiscoveryEnabled() {
+		t.Fatal("expected model discovery to be disabled")
+	}
+	if (&BackendConfig{}).ModelDiscoveryEnabled() == false {
+		t.Fatal("expected model discovery to default to enabled")
+	}
+}
+
 func TestRejectsInvalidUIRecentRequests(t *testing.T) {
 	cfg := Config{
 		Server: ServerConfig{Host: DefaultHost, Port: DefaultPort},

internal/proxy/models.go

@@ -26,6 +26,7 @@ type ModelDiscovery struct {
 	StatusCode int
 	Models     []DiscoveredModel
 	Err        error
+	Skipped    bool
 }
 
 // DiscoverModels queries every configured backend's OpenAI-compatible /models
@@ -36,6 +37,10 @@ func (p *Proxy) DiscoverModels(ctx context.Context) []ModelDiscovery {
 	results := make([]ModelDiscovery, len(p.backends))
 	var wg sync.WaitGroup
 	for i, backend := range p.backends {
+		if !backend.cfg.ModelDiscoveryEnabled() {
+			results[i] = ModelDiscovery{Backend: backend.cfg.Name, Skipped: true}
+			continue
+		}
 		wg.Add(1)
 		go func() {
 			defer wg.Done()
@@ -87,7 +92,7 @@ func (p *Proxy) discoverBackendModels(ctx context.Context, backend *resolvedBack
 		return result
 	}
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
-		result.Err = fmt.Errorf("HTTP %d: %s", resp.StatusCode, truncateForError(string(body), 200))
+		result.Err = fmt.Errorf("HTTP %d %s", resp.StatusCode, http.StatusText(resp.StatusCode))
 		return result
 	}
 
@@ -115,10 +120,3 @@ func (p *Proxy) discoverBackendModels(ctx context.Context, backend *resolvedBack
 	sort.Slice(result.Models, func(i, j int) bool { return result.Models[i].ID < result.Models[j].ID })
 	return result
 }
-
-func truncateForError(value string, limit int) string {
-	if len(value) <= limit {
-		return value
-	}
-	return value[:limit] + "..."
-}

internal/proxy/proxy_test.go

@@ -142,11 +142,30 @@ func TestDiscoverModelsReportsBackendHTTPError(t *testing.T) {
 	handler := mustNew(t, Options{Config: cfg, TokenProvider: StaticTokenProvider("token"), Metrics: NewMetrics()})
 
 	results := handler.DiscoverModels(context.Background())
-	if len(results) != 1 || results[0].Err == nil || !strings.Contains(results[0].Err.Error(), "HTTP 503") {
+	if len(results) != 1 || results[0].Err == nil || results[0].Err.Error() != "HTTP 503 Service Unavailable" {
 		t.Fatalf("expected per-backend HTTP error, got %#v", results)
 	}
 }
 
+func TestDiscoverModelsSkipsDisabledBackends(t *testing.T) {
+	called := false
+	upstream := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {
+		called = true
+	}))
+	defer upstream.Close()
+
+	disabled := false
+	cfg := testConfig([]config.Model{{ID: "model"}})
+	cfg.Backends[0].BaseURL = upstream.URL + "/v1"
+	cfg.Backends[0].ModelDiscovery = &disabled
+	handler := mustNew(t, Options{Config: cfg, TokenProvider: StaticTokenProvider("token"), Metrics: NewMetrics()})
+
+	results := handler.DiscoverModels(context.Background())
+	if len(results) != 1 || !results[0].Skipped || results[0].Err != nil || called {
+		t.Fatalf("expected disabled discovery without an upstream request, got %#v (called=%t)", results, called)
+	}
+}
+
 func TestChatCompletionsMissingModel(t *testing.T) {
 	handler := mustNew(t, Options{
 		Config:        testConfig([]config.Model{{ID: "google/gemini-2.5-flash"}}),

internal/ui/ui.go

@@ -438,7 +438,7 @@ func reconcileModels(cfg *config.Config, results []proxy.ModelDiscovery) []diagn
 				row.status, row.rowStyle = "MATCH", "green"
 				row.details = combinedModelDetails(backend, configured, found)
 				matched[upstreamID] = true
-			} else if !queried || result.Err != nil {
+			} else if !queried || result.Err != nil || result.Skipped {
 				row.status = "UNKNOWN"
 				row.rowStyle = "yellow"
 			}
@@ -659,7 +659,9 @@ func (m tuiModel) viewModels(tableWidth int) string {
 	}
 	if m.discoveryLoaded {
 		for _, result := range m.discoveryResults {
-			if result.Err != nil {
+			if result.Skipped {
+				fmt.Fprintf(&b, "%s %s\n", style(m.color, "muted").Render(result.Backend+":"), style(m.color, "muted").Render("model discovery disabled by configuration"))
+			} else if result.Err != nil {
 				fmt.Fprintf(&b, "%s %s\n", style(m.color, "red").Render(result.Backend+":"), style(m.color, "red").Render(result.Err.Error()))
 			} else {
 				fmt.Fprintf(&b, "%s %s\n", style(m.color, "green").Render(result.Backend+":"), style(m.color, "muted").Render(fmt.Sprintf("HTTP %d, %d models from %s", result.StatusCode, len(result.Models), result.URL)))

internal/ui/ui_test.go

@@ -98,6 +98,19 @@ func TestReconcileModelsUsesUnknownWhenDiscoveryFails(t *testing.T) {
 	}
 }
 
+func TestModelsViewShowsDisabledDiscoveryWithoutError(t *testing.T) {
+	m := testTUIModel()
+	m.activeTab = tabModels
+	m.discoveryLoaded = true
+	m.discoveryResults = []proxy.ModelDiscovery{{Backend: "google", Skipped: true}}
+	m.diagnosticModels = reconcileModels(m.cfg, m.discoveryResults)
+
+	view := m.viewModels(m.tableWidth())
+	if !strings.Contains(view, "model discovery disabled by configuration") {
+		t.Fatalf("expected disabled-discovery message, got:\n%s", view)
+	}
+}
+
 func TestModelsTabOpensDetailsAndReturns(t *testing.T) {
 	m := testTUIModel()
 	m.activeTab = tabModels