fix(model): register discovered endpoints with /v1 api_base

internal/model/discover.go

@@ -81,7 +81,11 @@ func buildDiscoveredProvider(ep inference.EndpointInfo) DiscoveredProvider {
 		if m.ID == "" {
 			continue
 		}
-		entries = append(entries, buildCustomEndpointEntry(m.ID, cluster, ""))
+		// The /v1 suffix is required — LiteLLM's OpenAI provider does not
+		// append it (CLAUDE.md pitfall 6). Without it, discovered vLLM
+		// endpoints 404 on /chat/completions and poison the model group
+		// alongside any correct `model setup custom` entry.
+		entries = append(entries, buildCustomEndpointEntry(m.ID, cluster+"/v1", ""))
 	}
 
 	return DiscoveredProvider{

internal/model/discover_test.go

@@ -99,8 +99,9 @@ func TestBuildDiscoveredProvider_TranslatesHost(t *testing.T) {
 	if first.LiteLLMParams.Model != "openai/meta-llama/Llama-3.1-8B-Instruct" {
 		t.Errorf("LiteLLMParams.Model = %q, want openai/meta-llama/Llama-3.1-8B-Instruct", first.LiteLLMParams.Model)
 	}
-	if first.LiteLLMParams.APIBase != "http://host.k3d.internal:8000" {
-		t.Errorf("APIBase = %q, want http://host.k3d.internal:8000", first.LiteLLMParams.APIBase)
+	// /v1 suffix required — LiteLLM's OpenAI provider does not append it.
+	if first.LiteLLMParams.APIBase != "http://host.k3d.internal:8000/v1" {
+		t.Errorf("APIBase = %q, want http://host.k3d.internal:8000/v1", first.LiteLLMParams.APIBase)
 	}
 	// buildCustomEndpointEntry sets api_key to "none" when no key is given.
 	if first.LiteLLMParams.APIKey != "none" {

internal/model/model.go

@@ -1574,6 +1574,12 @@ func buildCustomEndpointEntry(modelName, clusterEndpoint, apiKey string) ModelEn
 	return buildCustomEndpointEntryWithOptions(modelName, clusterEndpoint, apiKey, CustomEndpointOptions{})
 }
 
+// INVARIANT: clusterEndpoint must be the full OpenAI-compatible base
+// INCLUDING the /v1 suffix — LiteLLM's `openai/` provider sends requests
+// to <api_base>/chat/completions verbatim and never appends /v1
+// (CLAUDE.md pitfall 6). Both callers uphold this: AddCustomEndpoint
+// validates by probing <endpoint>/chat/completions, and discovery only
+// registers endpoints that answered <base>/v1/models (and appends /v1).
 func buildCustomEndpointEntryWithOptions(modelName, clusterEndpoint, apiKey string, options CustomEndpointOptions) ModelEntry {
 	entry := ModelEntry{
 		ModelName: modelName,
@@ -1696,20 +1702,6 @@ func decodeBase64(s string) (string, error) {
 	return string(decoded[:n]), nil
 }
 
-// WarnAndStripV1Suffix checks if an endpoint URL has a trailing /v1 suffix,
-// warns the user, and returns the stripped URL. For OpenAI-compatible providers,
-// LiteLLM auto-appends /v1, causing double /v1/v1 if the user includes it.
-func WarnAndStripV1Suffix(endpoint string) string {
-	trimmed := strings.TrimRight(endpoint, "/")
-	if strings.HasSuffix(trimmed, "/v1") {
-		fmt.Printf("  Warning: stripping trailing /v1 from endpoint URL (LiteLLM adds it automatically)\n")
-		fmt.Printf("  %s → %s\n", trimmed, strings.TrimSuffix(trimmed, "/v1"))
-
-		return strings.TrimSuffix(trimmed, "/v1")
-	}
-
-	return endpoint
-}
 
 // localhostToClusterEndpoint translates localhost URLs to k3d-internal URLs
 // so that services running on the host are reachable from inside the k3d cluster.