fix(runtime): improve local service reliability

Close SQLite handles safely on Windows, clean orphaned vectors, normalize Ollama endpoints, and surface model errors without leaking reasoning tags.

Co-authored-by: Cursor <cursoragent@cursor.com>

tests/test_smoke.py

@@ -217,16 +217,23 @@ def test_icon_source_requires_real_allowed_file_boundary(tmp_path):
 
 
 def test_ollama_stream_propagates_transport_errors(monkeypatch):
+    """A failing stream must surface, not be swallowed into an empty reply —
+    and it must carry Ollama's own explanation, since that is the only part the
+    user can act on. The response here is a real httpx.Response because the
+    error path reads the body, which a stubbed raise_for_status never exercised."""
+    import httpx
+
     class FailingResponse:
         async def __aenter__(self):
-            return self
+            return httpx.Response(
+                503,
+                json={"error": "Ollama unavailable"},
+                request=httpx.Request("POST", "http://127.0.0.1:11434/api/chat"),
+            )
 
         async def __aexit__(self, *args):
             return False
 
-        def raise_for_status(self):
-            raise RuntimeError("Ollama unavailable")
-
     class FailingClient:
         def __init__(self, **kwargs):
             pass
@@ -246,7 +253,7 @@ def test_ollama_stream_propagates_transport_errors(monkeypatch):
         async for _ in OllamaManager()._chat_stream([], "model", 0):
             pass
 
-    with pytest.raises(RuntimeError, match="Ollama unavailable"):
+    with pytest.raises(httpx.HTTPStatusError, match="Ollama unavailable"):
         import asyncio
         asyncio.run(consume())
 
@@ -264,13 +271,17 @@ def test_sync_compare_detects_direction(tmp_path):
     """The guard that stops a stale box from overwriting the other's chats.
     Both backup and restore refuse to run when this says the wrong thing, so a
     silent break here loses conversation history."""
+    import contextlib
     import sqlite3 as sq
     sync = _load_sync()
     db, dump = tmp_path / "memory.db", tmp_path / "memory.db.sql"
 
     def write(rows):
         db.unlink(missing_ok=True)
-        with sq.connect(db) as conn:
+        # closing() then the connection itself: sqlite3's own context manager
+        # commits but never closes, and Windows refuses to unlink a file that
+        # still has an open handle.
+        with contextlib.closing(sq.connect(db)) as conn, conn:
             conn.executescript(
                 # updated_at REAL, matching the production schema in store.py. A TEXT
                 # column here hid a real TypeError for months: the comparison in
@@ -482,3 +493,102 @@ def test_curator_drops_fabricated_facts():
     assert _reject_reason(
         "Jon prefers short answers over long explanations",
         "i really prefer short answers over long explanations") is None
+
+
+def test_ollama_failures_surface_the_reason_not_just_the_status():
+    """Ollama answers every failure with {"error": "..."} and httpx's default
+    message throws it away. A user hitting a retired cloud model saw
+    "Client error '410 Gone' for url ..." when the body said exactly why."""
+    import asyncio
+    import httpx
+    import pytest
+    from synapse.ollama_manager import _raise_for_ollama
+
+    req = httpx.Request("POST", "http://127.0.0.1:11434/api/chat")
+
+    retired = httpx.Response(410, json={"error": "glm-4.6 was retired at 2026-06-16"}, request=req)
+    with pytest.raises(httpx.HTTPStatusError) as ei:
+        asyncio.run(_raise_for_ollama(retired))
+    assert "retired" in str(ei.value) and "410" in str(ei.value)
+
+    # The common case, not just the exotic one.
+    missing = httpx.Response(404, json={"error": "model 'foo' not found"}, request=req)
+    with pytest.raises(httpx.HTTPStatusError) as ei:
+        asyncio.run(_raise_for_ollama(missing))
+    assert "model 'foo' not found" in str(ei.value)
+
+    # No usable body -> keep httpx's own wording rather than inventing one.
+    blank = httpx.Response(500, content=b"", request=req)
+    with pytest.raises(httpx.HTTPStatusError):
+        asyncio.run(_raise_for_ollama(blank))
+
+    # Success stays silent.
+    asyncio.run(_raise_for_ollama(httpx.Response(200, json={"ok": True}, request=req)))
+
+
+def test_ollama_host_is_normalized_for_clients_but_not_for_binding():
+    """OLLAMA_HOST is Ollama's *bind* variable, and `OLLAMA_HOST=0.0.0.0:11434`
+    is the normal way to expose it on a LAN. Used verbatim as a client base URL
+    it is unusable — no scheme, and 0.0.0.0 is not a destination — and every
+    request failed into list_models()'s bare `except: return []`, so the model
+    picker just went empty with no error anywhere."""
+    from synapse.nexus_config import _normalize_ollama_host as norm
+
+    assert norm("0.0.0.0:11434") == "http://127.0.0.1:11434"
+    assert norm("[::]:11434") == "http://127.0.0.1:11434"
+    assert norm("http://0.0.0.0:11434/") == "http://127.0.0.1:11434"
+    assert norm("127.0.0.1:11434") == "http://127.0.0.1:11434"   # scheme supplied
+    assert norm("") == "http://127.0.0.1:11434"
+    # A real remote is deliberate — leave it alone.
+    assert norm("https://ollama.lan:11434") == "https://ollama.lan:11434"
+    assert norm("192.168.1.50:11434") == "http://192.168.1.50:11434"
+
+
+def test_spawned_serve_keeps_the_users_bind_address(monkeypatch):
+    """Normalizing for the client must not quietly un-expose a server we spawn."""
+    import importlib
+    from synapse import nexus_config
+    monkeypatch.setenv("OLLAMA_HOST", "0.0.0.0:11434")
+    reloaded = importlib.reload(nexus_config)
+    try:
+        assert reloaded.settings.ollama_bind == "0.0.0.0:11434"      # listens everywhere
+        assert reloaded.settings.ollama_host == "http://127.0.0.1:11434"  # we connect here
+    finally:
+        monkeypatch.delenv("OLLAMA_HOST", raising=False)
+        importlib.reload(nexus_config)
+
+
+def test_think_blocks_never_reach_the_reply():
+    """A reasoning model emits <think> inline in `content` even with think off,
+    and the whole internal monologue reached the chat window — including the
+    literal closing tags. Streaming has to cope with a tag split across chunks,
+    and with the shape actually observed: a stray </think> and no opening."""
+    from synapse.ollama_manager import ThinkStripper, strip_think
+
+    def stream(chunks):
+        s = ThinkStripper()
+        return "".join(s.feed(c) for c in chunks) + s.flush()
+
+    assert stream(["hello ", "<think>", "noise", "</think>", "world"]) == "hello world"
+    # tag split across chunk boundaries
+    assert stream(["a<th", "ink>x</thi", "nk>b"]) == "ab"
+    # never closed -> it was all reasoning
+    assert stream(["keep", "<think>", "runs off the end"]) == "keep"
+    # stray close, no open: at minimum the tag itself must not be shown
+    assert "</think>" not in stream(["reasoning...", "</think>", "the answer"])
+    # ordinary text is untouched, including angle brackets
+    assert stream(["a < b ", "and c > d"]) == "a < b and c > d"
+
+    # With the complete message the stray-close case can be handled properly:
+    # everything before it was reasoning.
+    assert strip_think("rambling\n</think>\nThe answer") == "The answer"
+    assert strip_think("a<think>b</think>c") == "ac"
+    assert strip_think("no tags here") == "no tags here"
+
+
+def test_render_hint_is_only_added_when_the_tool_is_offered():
+    """The hint and the tool must share one condition. Unconditional, it turned
+    up recited as fact inside an answer about LRU caches."""
+    from synapse import tools as t
+    assert t.wants_render_preview("draw me a chart")
+    assert not t.wants_render_preview("Summarize what a thread-safe LRU cache needs")