feat: switch LLM backend to OpenAI-compat; add cf-orch remote inference support

Turnstone now calls /v1/chat/completions instead of Ollama's /api/generate.
This format works with both local Ollama (>=0.1.24) and a remote cf-orch
coordinator, enabling GPU-less nodes like Xander's to route diagnoses through
the cluster without any local model.

- llm.py: OpenAI-compat messages format, optional Bearer auth header
- diagnose.py: thread llm_api_key through the call chain
- rest.py: llm_api_key pref (default empty), SettingsBody field, passed to diagnose
- SettingsView.vue: API Key field, label updated from "Ollama URL" to "LLM Endpoint URL"
- tests: updated mocks for new response shape; added bearer token assertion test

app/rest.py

@@ -79,6 +79,7 @@ _PREFS_DEFAULTS: dict = {
     "entry_point_style": "topbar",
     "llm_url": "http://localhost:11434",
     "llm_model": "llama3.1:8b",
+    "llm_api_key": "",
     "severity_overrides": [
         {
             "name": "PAM auth noise",
@@ -121,6 +122,7 @@ class SettingsBody(BaseModel):
     entry_point_style: str | None = None
     llm_url: str | None = None
     llm_model: str | None = None
+    llm_api_key: str | None = None
     severity_overrides: list[SeverityOverride] | None = None
 
 
@@ -249,6 +251,7 @@ def diagnose_post(body: DiagnoseRequest) -> dict:
         until=body.until,
         llm_url=prefs.get("llm_url") or None,
         llm_model=prefs.get("llm_model") or None,
+        llm_api_key=prefs.get("llm_api_key") or None,
     )
     return {
         "summary": result["summary"],
@@ -273,6 +276,8 @@ def patch_settings(body: SettingsBody) -> dict:
         prefs["llm_url"] = body.llm_url
     if body.llm_model is not None:
         prefs["llm_model"] = body.llm_model
+    if body.llm_api_key is not None:
+        prefs["llm_api_key"] = body.llm_api_key
     if body.severity_overrides is not None:
         prefs["severity_overrides"] = [o.model_dump() for o in body.severity_overrides]
     _save_prefs(prefs)

app/services/diagnose.py

@@ -51,6 +51,7 @@ def diagnose(
     until: str | None = None,
     llm_url: str | None = None,
     llm_model: str | None = None,
+    llm_api_key: str | None = None,
 ) -> dict[str, Any]:
     """Run layered log search with NL time extraction. Returns summary + entries."""
     time_detected = since is not None and until is not None
@@ -84,7 +85,7 @@ def diagnose(
 
     reasoning: str | None = None
     if llm_url and llm_model:
-        reasoning = summarize(query, combined, llm_url=llm_url, llm_model=llm_model)
+        reasoning = summarize(query, combined, llm_url=llm_url, llm_model=llm_model, api_key=llm_api_key)
 
     return {
         "summary": {

app/services/llm.py

@@ -37,20 +37,30 @@ def summarize(
     entries: list[SearchResult],
     llm_url: str,
     llm_model: str,
+    api_key: str | None = None,
     timeout: float = 20.0,
 ) -> str | None:
     if not entries:
         return None
     log_block = _build_context(entries)
     prompt = _PROMPT_TEMPLATE.format(query=query, n=min(len(entries), 25), log_block=log_block)
+    headers = {"Authorization": f"Bearer {api_key}"} if api_key else {}
     try:
         resp = httpx.post(
-            f"{llm_url.rstrip('/')}/api/generate",
+            f"{llm_url.rstrip('/')}/v1/chat/completions",
-            json={"model": llm_model, "prompt": prompt, "stream": False},
+            json={
+                "model": llm_model,
+                "messages": [{"role": "user", "content": prompt}],
+                "stream": False,
+            },
+            headers=headers,
             timeout=timeout,
         )
         resp.raise_for_status()
-        return resp.json().get("response", "").strip() or None
+        choices = resp.json().get("choices") or []
+        if not choices:
+            return None
+        return (choices[0].get("message", {}).get("content") or "").strip() or None
     except Exception as exc:
         logger.warning("LLM summarization failed (%s): %s", type(exc).__name__, exc)
         return None

tests/test_services_llm.py

@@ -39,7 +39,16 @@ def test_summarize_returns_none_on_http_error():
 def test_summarize_returns_none_on_empty_response():
     mock_resp = MagicMock()
     mock_resp.raise_for_status.return_value = None
-    mock_resp.json.return_value = {"response": ""}
+    mock_resp.json.return_value = {"choices": [{"message": {"content": ""}}]}
+    with patch("app.services.llm.httpx.post", return_value=mock_resp):
+        result = summarize("query", [_entry("x")], "http://host", "llama3")
+    assert result is None
+
+
+def test_summarize_returns_none_on_missing_choices():
+    mock_resp = MagicMock()
+    mock_resp.raise_for_status.return_value = None
+    mock_resp.json.return_value = {"choices": []}
     with patch("app.services.llm.httpx.post", return_value=mock_resp):
         result = summarize("query", [_entry("x")], "http://host", "llama3")
     assert result is None
@@ -48,12 +57,22 @@ def test_summarize_returns_none_on_empty_response():
 def test_summarize_returns_text_on_success():
     mock_resp = MagicMock()
     mock_resp.raise_for_status.return_value = None
-    mock_resp.json.return_value = {"response": "Ollama exited with code 1."}
+    mock_resp.json.return_value = {"choices": [{"message": {"content": "Ollama exited with code 1."}}]}
     with patch("app.services.llm.httpx.post", return_value=mock_resp):
         result = summarize("ollama crashed", [_entry("Failed")], "http://host", "llama3")
     assert result == "Ollama exited with code 1."
 
 
+def test_summarize_sends_bearer_token():
+    mock_resp = MagicMock()
+    mock_resp.raise_for_status.return_value = None
+    mock_resp.json.return_value = {"choices": [{"message": {"content": "disk full"}}]}
+    with patch("app.services.llm.httpx.post", return_value=mock_resp) as mock_post:
+        summarize("disk error", [_entry("ENOSPC")], "http://host", "llama3", api_key="test-key")
+    call_kwargs = mock_post.call_args
+    assert call_kwargs.kwargs["headers"] == {"Authorization": "Bearer test-key"}
+
+
 def test_build_context_sorts_errors_first():
     entries = [
         _entry("info message", severity="INFO"),

web/src/views/SettingsView.vue

@@ -36,11 +36,13 @@
       <div>
         <h2 class="text-text-primary text-sm font-semibold mb-1">LLM Reasoning</h2>
         <p class="text-text-dim text-xs mb-3">
-          Ollama endpoint used to generate plain-language diagnoses. Leave blank to disable.
+          LLM endpoint for plain-language diagnoses. Works with local Ollama or a remote
+          cf-orch coordinator (e.g. <span class="font-mono">https://orch.circuitforge.tech</span>).
+          Leave blank to disable.
         </p>
         <div class="space-y-3">
           <div>
-            <label class="block text-xs text-text-dim mb-1">Ollama URL</label>
+            <label class="block text-xs text-text-dim mb-1">LLM Endpoint URL</label>
             <input
               v-model="prefs.llm_url"
               type="text"
@@ -57,6 +59,18 @@
               class="w-full bg-surface border border-surface-border rounded px-3 py-2 text-sm text-text-primary placeholder-text-dim focus:outline-none focus:border-accent transition-colors"
             />
           </div>
+          <div>
+            <label class="block text-xs text-text-dim mb-1">
+              API Key
+              <span class="text-text-dim font-normal">(optional — required for cf-orch remote inference)</span>
+            </label>
+            <input
+              v-model="prefs.llm_api_key"
+              type="password"
+              placeholder="Leave blank for local Ollama"
+              class="w-full bg-surface border border-surface-border rounded px-3 py-2 text-sm text-text-primary placeholder-text-dim focus:outline-none focus:border-accent transition-colors"
+            />
+          </div>
           <button
             @click="saveLlm"
             class="px-4 py-2 bg-accent text-surface text-sm rounded font-medium hover:opacity-90 transition-opacity"
@@ -171,10 +185,11 @@ interface Prefs {
   entry_point_style: 'topbar' | 'fab'
   llm_url: string
   llm_model: string
+  llm_api_key: string
   severity_overrides: SeverityOverride[]
 }
 
-const prefs      = ref<Prefs>({ entry_point_style: 'topbar', llm_url: '', llm_model: '', severity_overrides: [] })
+const prefs      = ref<Prefs>({ entry_point_style: 'topbar', llm_url: '', llm_model: '', llm_api_key: '', severity_overrides: [] })
 const saveStatus = ref<{ ok: boolean; msg: string } | null>(null)
 const showAddOverride = ref(false)
 const newRule = ref<SeverityOverride>({ name: '', pattern: '', override_severity: 'WARN', enabled: true })
@@ -215,7 +230,7 @@ async function setEntryPoint(style: 'topbar' | 'fab') {
 async function saveLlm() {
   saveStatus.value = null
   try {
-    await patch({ llm_url: prefs.value.llm_url, llm_model: prefs.value.llm_model })
+    await patch({ llm_url: prefs.value.llm_url, llm_model: prefs.value.llm_model, llm_api_key: prefs.value.llm_api_key })
     saveStatus.value = { ok: true, msg: 'LLM settings saved' }
     setTimeout(() => { saveStatus.value = null }, 2000)
   } catch {