fix(inference): probe /v1/models for OpenAI-compatible local runtimes (Closes #5053)

src/openhuman/inference/local/provider.rs

@@ -48,6 +48,64 @@ pub(crate) fn provider_from_config(config: &Config) -> LocalAiProvider {
     }
 }
 
+/// How a local runtime exposes its installed-model catalog.
+///
+/// Ollama serves a native `GET /api/tags` listing; every OpenAI-compatible
+/// runtime (LM Studio, OMLX, `local-openai`, and any custom BYOK endpoint that
+/// speaks the OpenAI `/v1` surface) serves `GET /v1/models`. Sending an Ollama
+/// probe to an OpenAI-compatible server produces `GET /v1/api/tags`, which
+/// LM Studio logs as `Unexpected endpoint or method` and answers with an empty
+/// catalog — so model discovery silently fails and the model never appears as
+/// selectable (GH #5053).
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) enum ModelDiscoveryApi {
+    /// Ollama-native `/api/tags`.
+    OllamaTags,
+    /// OpenAI-compatible `/v1/models`.
+    OpenAiModels,
+}
+
+/// True when `base_url`'s path is the OpenAI-compatible `/v1` root — the
+/// canonical marker of an OpenAI-style server (LM Studio, OMLX, LiteLLM, …).
+///
+/// A genuine Ollama base URL is host-rooted (`http://localhost:11434`) with no
+/// `/v1` path segment, so this cleanly separates the two API shapes by
+/// **endpoint type**, not by "is it localhost" (which would misidentify a
+/// localhost LM Studio server as Ollama — the #5053 conflation).
+pub(crate) fn endpoint_is_openai_v1(base_url: &str) -> bool {
+    if let Ok(url) = reqwest::Url::parse(base_url.trim()) {
+        let path = url.path().trim_end_matches('/').to_ascii_lowercase();
+        return path.ends_with("/v1");
+    }
+    // Fall back to a lexical check when the URL doesn't parse cleanly.
+    base_url
+        .trim()
+        .trim_end_matches('/')
+        .to_ascii_lowercase()
+        .ends_with("/v1")
+}
+
+/// Select the model-discovery API for a local runtime from its provider slug
+/// and base URL — by provider **type**, never by "is it localhost".
+///
+/// Genuine Ollama (`provider = "ollama"` on a host-rooted base) uses
+/// `/api/tags`. Every OpenAI-compatible runtime uses `/v1/models`: an explicit
+/// `lm_studio` / `omlx` slug, OR any endpoint whose path is the OpenAI `/v1`
+/// root. The `/v1` endpoint clause is what rescues a custom BYOK localhost
+/// endpoint (e.g. LM Studio on `http://localhost:1234/v1`) whose provider tag
+/// still defaults to `ollama`: the endpoint type wins over the fallback slug
+/// (GH #5053).
+pub(crate) fn model_discovery_api(provider: &str, base_url: &str) -> ModelDiscoveryApi {
+    if endpoint_is_openai_v1(base_url) {
+        return ModelDiscoveryApi::OpenAiModels;
+    }
+    match normalize_provider(provider).as_str() {
+        "ollama" => ModelDiscoveryApi::OllamaTags,
+        // lm_studio, omlx, and any other OpenAI-compatible local runtime.
+        _ => ModelDiscoveryApi::OpenAiModels,
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -71,4 +129,53 @@ mod tests {
         assert_eq!(normalize_provider("omlx-server"), "omlx");
         assert_eq!(normalize_provider("OMLX"), "omlx");
     }
+
+    #[test]
+    fn endpoint_is_openai_v1_detects_v1_root() {
+        assert!(endpoint_is_openai_v1("http://localhost:1234/v1"));
+        assert!(endpoint_is_openai_v1("http://localhost:1234/v1/"));
+        assert!(endpoint_is_openai_v1("https://box.local:1234/openai/v1"));
+        // Genuine Ollama base is host-rooted with no /v1 path.
+        assert!(!endpoint_is_openai_v1("http://localhost:11434"));
+        assert!(!endpoint_is_openai_v1("http://localhost:11434/"));
+        // A `/v1` embedded mid-path is not the OpenAI root.
+        assert!(!endpoint_is_openai_v1("http://localhost:11434/v1/models"));
+    }
+
+    #[test]
+    fn model_discovery_api_uses_tags_for_genuine_ollama() {
+        // Ollama slug on its host-rooted native base -> /api/tags.
+        assert_eq!(
+            model_discovery_api("ollama", "http://localhost:11434"),
+            ModelDiscoveryApi::OllamaTags
+        );
+        assert_eq!(
+            model_discovery_api("", "http://localhost:11434"),
+            ModelDiscoveryApi::OllamaTags
+        );
+    }
+
+    #[test]
+    fn model_discovery_api_uses_v1_models_for_openai_compatible() {
+        // Explicit LM Studio / OMLX slugs are OpenAI-compatible.
+        assert_eq!(
+            model_discovery_api("lm_studio", "http://localhost:1234/v1"),
+            ModelDiscoveryApi::OpenAiModels
+        );
+        assert_eq!(
+            model_discovery_api("omlx", "http://localhost:8080/v1"),
+            ModelDiscoveryApi::OpenAiModels
+        );
+        // The #5053 case: a custom BYOK OpenAI-compatible endpoint on localhost
+        // whose provider tag still defaults to `ollama` must NOT be probed with
+        // /api/tags — the `/v1` endpoint type wins.
+        assert_eq!(
+            model_discovery_api("ollama", "http://localhost:1234/v1"),
+            ModelDiscoveryApi::OpenAiModels
+        );
+        assert_eq!(
+            model_discovery_api("custom-byok", "http://localhost:1234/v1"),
+            ModelDiscoveryApi::OpenAiModels
+        );
+    }
 }

src/openhuman/inference/local/service/ollama_admin/diagnostics.rs

@@ -7,7 +7,9 @@ use crate::openhuman::inference::local::ollama::{
     ollama_base_url_from_config, OllamaModelShow, OllamaModelTag, OllamaShowRequest,
     OllamaShowResponse, OllamaTagsResponse,
 };
-use crate::openhuman::inference::local::provider::{provider_from_config, LocalAiProvider};
+use crate::openhuman::inference::local::provider::{
+    model_discovery_api, provider_from_config, LocalAiProvider, ModelDiscoveryApi,
+};
 use crate::openhuman::inference::model_ids;
 use crate::openhuman::inference::presets::{self, VisionMode};
 
@@ -23,6 +25,26 @@ impl LocalAiService {
         }
 
         let base_url = ollama_base_url_from_config(config);
+
+        // Route OpenAI-compatible local runtimes to the `/v1/models` probe.
+        // `provider_from_config` collapses everything that is not literally
+        // `lm_studio` onto `Ollama`, so an OMLX / `local-openai` / custom BYOK
+        // endpoint on the OpenAI `/v1` surface (e.g. LM Studio at
+        // `http://localhost:1234/v1`) would otherwise be probed at
+        // `<base>/api/tags` -> `GET /v1/api/tags`, which LM Studio rejects as an
+        // unexpected endpoint and answers with an empty catalog — silently
+        // breaking model discovery (GH #5053). Decide by endpoint *type*, never
+        // by "is it localhost".
+        if model_discovery_api(&config.local_ai.provider, &base_url)
+            == ModelDiscoveryApi::OpenAiModels
+        {
+            log::debug!(
+                "[local_ai] diagnostics: base_url={} is OpenAI-compatible (/v1) — routing to /v1/models discovery instead of Ollama /api/tags",
+                base_url
+            );
+            return self.lm_studio_diagnostics(config).await;
+        }
+
         let healthy = self.ollama_healthy_at(&base_url).await;
         let runner_ok = if healthy {
             self.ollama_runner_ok_at(&base_url).await

src/openhuman/inference/local/service/ollama_admin_tests.rs

@@ -757,6 +757,58 @@ async fn lm_studio_diagnostics_reports_loaded_chat_model() {
     assert_eq!(diag["ok"], true);
 }
 
+/// Regression for GH #5053: a custom OpenAI-compatible BYOK endpoint on
+/// localhost (e.g. LM Studio at `http://localhost:1234/v1`) whose `provider`
+/// tag still defaults to `ollama` must be probed with `/v1/models`, NOT the
+/// Ollama-native `/api/tags`. The mock serves ONLY `/v1/models` and no
+/// `/api/tags`, so before the fix diagnostics took the Ollama branch,
+/// hit an unrouted `/v1/api/tags`, and reported the model absent; after the
+/// fix the `/v1` endpoint type routes discovery to `/v1/models` and the model
+/// is found.
+#[tokio::test]
+async fn diagnostics_openai_compatible_v1_endpoint_uses_v1_models_not_api_tags() {
+    let _guard = crate::openhuman::inference::inference_test_guard();
+
+    // OpenAI-compatible server: exposes `/v1/models` and deliberately no
+    // `/api/tags` — an Ollama probe here would 404 (silently empty discovery).
+    let app = Router::new().route(
+        "/v1/models",
+        get(|| async {
+            Json(json!({
+                "data": [
+                    { "id": "local-model", "object": "model", "owned_by": "lm-studio" }
+                ]
+            }))
+        }),
+    );
+    let base = spawn_mock(app).await;
+
+    // The #5053 config: a `/v1` OpenAI-compatible endpoint whose provider tag is
+    // the defaulted `ollama` (not `lm_studio`).
+    let mut config = lm_studio_config(&base);
+    config.local_ai.provider = "ollama".to_string();
+
+    let service = LocalAiService::new(&config);
+    let diag = service.diagnostics(&config).await.expect("diagnostics");
+
+    // `lm_studio_running` is emitted only by the OpenAI-compatible (`/v1/models`)
+    // diagnostics path — the Ollama branch reports `ollama_running` and leaves
+    // this key null. Its presence proves discovery was routed by endpoint type,
+    // not sent to `/api/tags`.
+    assert_eq!(diag["lm_studio_running"], true);
+    let installed = diag["installed_models"]
+        .as_array()
+        .cloned()
+        .unwrap_or_default();
+    assert!(
+        installed
+            .iter()
+            .any(|m| m["name"].as_str() == Some("local-model")),
+        "OpenAI-compatible /v1 endpoint must discover models via /v1/models, got: {:?}",
+        installed
+    );
+}
+
 #[tokio::test]
 async fn lm_studio_diagnostics_flags_missing_chat_model() {
     let _guard = crate::openhuman::inference::inference_test_guard();