fix: recover empty validation compilations

dashboard/manifest.json

@@ -3,7 +3,7 @@
   "label": "Ollama Models",
   "description": "Inspect, manage, and chat with local Ollama models, including shared persistent conversations, performance metrics, images, PDFs, URLs, and live memory telemetry.",
   "icon": "Cpu",
-  "version": "1.7.9",
+  "version": "1.7.10",
   "tab": {"path": "/ollama-manager", "position": "after:models"},
   "entry": "dist/index.js",
   "css": "dist/style.css",

dashboard/plugin_api.py

@@ -2150,13 +2150,13 @@ def _run_validation_harness(
     metrics.append(_persist_metric(conversation_id, final_id, primary, final_state, status="completed"))
     final_message = final_result.get("message") if isinstance(final_result.get("message"), dict) else {}
     final_content = str(final_message.get("content") or "").strip()
-    if not final_content:
-        raise HTTPException(502, "Primary model returned an empty compiled answer")
-    if _looks_like_validation_report(final_content):
+    compiler_needs_retry = not final_content or _looks_like_validation_report(final_content)
+    if compiler_needs_retry:
         retry_id = uuid.uuid4().hex
         retry_started = time.time()
+        retry_reason = "empty output" if not final_content else "review text"
         _persist_chat_stage(request_id, "compiler-retry", primary, "running", retry_started)
-        _persist_chat_event(request_id, conversation_id, "compiler-retry-started", level="warning", stage="Primary returned review text; requesting final answer", model=primary)
+        _persist_chat_event(request_id, conversation_id, "compiler-retry-started", level="warning", stage=f"Primary returned {retry_reason}; requesting final answer", model=primary)
         _chat_state(retry_id, state="preparing", stage="Preparing final-answer retry", model=primary, parent_id=request_id, conversation_id=conversation_id)
         retry_payload = _chat_payload(
             body,
@@ -2178,8 +2178,8 @@ def _run_validation_harness(
             _persist_chat_event(request_id, conversation_id, "compiler-retry-completed", stage="Final answer retry completed", model=primary, payload={"output_chars": len(final_content)})
         else:
             final_content = draft
-            _persist_chat_stage(request_id, "compiler-retry", primary, "fallback", retry_started, finished_at=retry_finished, output_chars=len(final_content), error="Primary returned review text twice; preserved the primary draft")
-            _persist_chat_event(request_id, conversation_id, "compiler-retry-fallback", level="warning", stage="Preserved primary draft after invalid finalization", model=primary, payload={"output_chars": len(final_content)})
+            _persist_chat_stage(request_id, "compiler-retry", primary, "fallback", retry_started, finished_at=retry_finished, output_chars=len(final_content), error=f"Primary returned {retry_reason} twice; preserved the primary draft")
+            _persist_chat_event(request_id, conversation_id, "compiler-retry-fallback", level="warning", stage="Preserved primary draft after invalid finalization", model=primary, payload={"output_chars": len(final_content), "reason": retry_reason})
     compiler_finished = time.time()
     _persist_chat_stage(request_id, "compiler", primary, "completed", compiler_started, finished_at=compiler_finished, output_chars=len(final_content))
     _persist_chat_event(request_id, conversation_id, "compiler-completed", stage="Primary final answer compiled", model=primary, payload={"output_chars": len(final_content)})

plugin.yaml

@@ -1,5 +1,5 @@
 name: ollama-manager
-version: 1.7.9
+version: 1.7.10
 description: Native dashboard manager and chat interface for local Ollama models, attachments, URLs, shared persistent conversations, performance metrics, and live runtime telemetry.
 auto_install_dependencies: true
 python_dependencies:

tests/test_validation_harness.py

@@ -178,6 +178,42 @@ class ValidationHarnessTests(unittest.TestCase):
         self.assertEqual(len(metrics), 4)
         self.assertTrue(any("IMPORTANT FINALIZATION RETRY" in prompt for _, prompt in calls))
 
+    def test_empty_compiler_output_is_retried_and_never_becomes_a_502(self):
+        body = api.ChatRequest(primary_model="primary", validator_models=["validator"], harness=True, message="Write the requested result")
+        calls = []
+
+        def fake_stream(payload, request_id, cancel_event=None, parent_id=None):
+            model = payload["model"]
+            prompt = payload["messages"][-1]["content"]
+            calls.append((model, prompt))
+            if model == "validator":
+                content = "Make the result more specific."
+            elif "IMPORTANT FINALIZATION RETRY" in prompt:
+                content = "refined final answer after empty compiler output"
+            elif "VALIDATION REPORTS:" in prompt:
+                content = ""
+            else:
+                content = "primary draft"
+            return {"message": {"role": "assistant", "content": content}, "done": True}
+
+        def fake_metric(conversation_id, request_id, model, state, status=None, error=""):
+            return {"request_id": request_id, "model": model, "status": status}
+
+        with patch.object(api, "_require_installed_model", side_effect=lambda name: name), patch.object(
+            api, "_stream_chat_request", side_effect=fake_stream
+        ), patch.object(api, "_persist_metric", side_effect=fake_metric), patch.object(
+            api, "_persist_chat_stage"
+        ), patch.object(api, "_persist_chat_event"), patch.object(api, "_chat_state"):
+            final, metrics, reports = api._run_validation_harness(
+                body, "root-request", "conversation", "primary", ["validator"], threading.Event(), []
+            )
+
+        self.assertEqual(final, "refined final answer after empty compiler output")
+        self.assertEqual(len(reports), 1)
+        self.assertEqual(len(metrics), 4)
+        self.assertTrue(any("Primary returned empty output" in prompt for _, prompt in calls) or any("IMPORTANT FINALIZATION RETRY" in prompt for _, prompt in calls))
+
+
     def test_status_omits_full_catalog_by_default(self):
         with patch.object(api, "_local_tags", return_value=[]), patch.object(api, "_local_ps", return_value=[]), patch.object(
             api, "_ensure_catalog", return_value={"models": [], "families": {}, "source": "test"}