diff --git a/src/fabledassistant/services/llm.py b/src/fabledassistant/services/llm.py
index 288df68..a01edc6 100644
--- a/src/fabledassistant/services/llm.py
+++ b/src/fabledassistant/services/llm.py
@@ -159,6 +159,24 @@ async def wait_for_model_loaded(model: str, timeout: float = 90.0) -> bool:
         await asyncio.sleep(min(2.0, remaining))
 
 
+async def _raise_ollama_error(resp: httpx.Response, model: str) -> None:
+    """On a non-2xx Ollama response, log its body before raising.
+
+    ``resp.raise_for_status()`` alone throws away Ollama's error body, which
+    carries the actual reason (e.g. ``"model does not support tools"``,
+    ``"context length exceeded"``). Reading the body first means failures
+    surface a usable message instead of a bare HTTPStatusError.
+    """
+    if resp.status_code < 400:
+        return
+    body = (await resp.aread()).decode("utf-8", errors="replace").strip()
+    logger.error(
+        "Ollama /api/chat %d for model=%s: %s",
+        resp.status_code, model, body[:500],
+    )
+    resp.raise_for_status()
+
+
 async def stream_chat(
     messages: list[dict],
     model: str,
@@ -184,7 +202,7 @@ async def stream_chat(
             f"{Config.OLLAMA_URL}/api/chat",
             json=payload,
         ) as resp:
-            resp.raise_for_status()
+            await _raise_ollama_error(resp, model)
             async for line in resp.aiter_lines():
                 if not line.strip():
                     continue
@@ -245,7 +263,7 @@ async def stream_chat_with_tools(
             f"{Config.OLLAMA_URL}/api/chat",
             json=payload,
         ) as resp:
-            resp.raise_for_status()
+            await _raise_ollama_error(resp, model)
             accumulated_tool_calls: list[dict] = []
             async for line in resp.aiter_lines():
                 if not line.strip():
diff --git a/src/fabledassistant/services/projects.py b/src/fabledassistant/services/projects.py
index d3e1e38..16080a4 100644
--- a/src/fabledassistant/services/projects.py
+++ b/src/fabledassistant/services/projects.py
@@ -141,13 +141,30 @@ async def generate_project_summary(user_id: int, project_id: int) -> None:
         logger.debug("Failed to generate summary for project %d", project_id, exc_info=True)
 
 
+# Cutoff: summaries older than this were generated by qwen2.5:3b, which
+# produced broken output (token repetition, hallucinated topics, misspellings).
+# Anything stamped before the switch to gemma3:4b on 2026-04-12 gets
+# regenerated at startup so the whole project list ends up on the better model.
+_BACKGROUND_MODEL_CUTOVER = datetime(2026, 4, 12, tzinfo=timezone.utc)
+
+
 async def backfill_project_summaries() -> None:
-    """Generate summaries for all projects missing auto_summary. Fire-and-forget."""
+    """Generate summaries for projects missing or predating the gemma3:4b cutover.
+
+    Fire-and-forget: each summary runs as its own background task.
+    """
     import asyncio
+    from sqlalchemy import or_
     try:
         async with async_session() as session:
             rows = (await session.execute(
-                select(Project.id, Project.user_id).where(Project.auto_summary.is_(None))
+                select(Project.id, Project.user_id).where(
+                    or_(
+                        Project.auto_summary.is_(None),
+                        Project.summary_updated_at.is_(None),
+                        Project.summary_updated_at < _BACKGROUND_MODEL_CUTOVER,
+                    )
+                )
             )).all()
         for row in rows:
             asyncio.create_task(generate_project_summary(row.user_id, row.id))