feat: model-written titles, fix empty replies after search

Titles: once a reply lands in a session still wearing its placeholder
title, the exchange goes back to the model in a short side request asking
for six words or fewer. Keyed on the placeholder rather than the turn
number, so an empty first reply does not forfeit titling for the session
and a one-shot window is titled from the question that was answered.
Thinking is disabled for that request, or a reasoning model spends the
whole budget thinking and returns nothing.

Empty replies: the final search round now says so in the tool result.
Withdrawing the tool schema is invisible to the model, which asks for
another search regardless; the request then surfaces as literal
<tool_call> text or vanishes into the thinking block, leaving the reply
empty either way. The note rides on the tool result because Qwen3.5's
template rejects a trailing system message outright.

max_searches now defaults to 2. The other half of that bug was
llama-server's quantized KV cache: with cache-type-k/v = q8_0 three of
five turns broke, and f16 answered six of six.

Measured against a live router, not mocked.

1 files changed, 156 insertions, 5 deletions

diff --git a/test_llamachat.py b/test_llamachat.py
index a766382..a06d737 100755
--- a/test_llamachat.py
+++ b/test_llamachat.py
@@ -967,11 +967,11 @@ def test_search_tool_schema():
     assert out == [("content", "hi")]
     assert sent == [None], sent
 
-    # One search per turn by default: a follow-up call arrives as literal
-    # <tool_call> XML rather than a structured call, and a turn that ends on
-    # one shows the user an empty reply.
-    assert config.DEFAULTS["max_searches"] == 1
-    assert config.load(Path("/nonexistent/config.toml")).max_searches == 1
+    # Two searches per turn, which a follow-up call now survives: the final
+    # round's tool result says the tool is gone, and quantized KV cache, the
+    # other half of the empty-reply bug, is a server-side setting.
+    assert config.DEFAULTS["max_searches"] == 2
+    assert config.load(Path("/nonexistent/config.toml")).max_searches == 2
 
     # A configured-but-urlless setup resolves to disabled at config load.
     with tempfile.TemporaryDirectory() as tmp:
@@ -1472,6 +1472,152 @@ def test_search_html():
     print("ok  search html")
 
 
+def test_final_round_note():
+    """The last search result tells the model it has no searches left.
+
+    Withdrawing the tool schema is invisible to the model: it tries a second
+    search anyway and emits it as literal <tool_call> text, or trails off
+    inside the thinking block, either way leaving the reply empty. Saying so
+    in the tool result is what actually ends the turn with an answer. The
+    note cannot be a trailing system message: this model's chat template
+    raises "System message must be at the beginning".
+    """
+    from llamachat import search
+
+    plain = search.tool_message("call_1", [{"title": "T", "url": "u", "content": "c"}])
+    final = search.tool_message(
+        "call_1", [{"title": "T", "url": "u", "content": "c"}], last=True
+    )
+    assert plain["role"] == "tool" and final["role"] == "tool"
+    assert search.NO_MORE_SEARCHES not in plain["content"]
+    assert search.NO_MORE_SEARCHES in final["content"]
+    # The results themselves survive the note being appended.
+    for message in (plain, final):
+        assert '"title": "T"' in message["content"] or '"title":"T"' in message["content"]
+    # A failed final search still gets the note, or the model retries.
+    failed = search.tool_message("call_1", None, error="boom", last=True)
+    assert "boom" in failed["content"]
+    assert search.NO_MORE_SEARCHES in failed["content"]
+    print("ok  final round note")
+
+
+def test_title_cleaning():
+    clean = backend.clean_title
+
+    assert clean("Kernel build failure") == "Kernel build failure"
+    # Quotes, trailing punctuation and a leading label all come off.
+    assert clean('"Kernel build failure"') == "Kernel build failure"
+    assert clean("Title: Kernel build failure.") == "Kernel build failure"
+    assert clean("**Kernel build failure**") == "Kernel build failure"
+    # Thinking that leaks into the body is dropped, answer kept.
+    assert clean("<think>hmm, short</think>Kernel build") == "Kernel build"
+    # An unclosed think block leaves nothing usable rather than a stray tag.
+    assert clean("<think>still reasoning") == ""
+    # Chatter before the title: the last non-empty line wins.
+    assert clean("Sure, here you go:\nKernel build failure") == "Kernel build failure"
+    # Too long is cut on a word boundary, not mid-word.
+    long = clean("word " * 40)
+    assert len(long) <= 60, long
+    assert not long.endswith("wor"), long
+    # Nothing usable yields nothing, so the caller keeps its fallback.
+    assert clean("") == ""
+    assert clean("   \n\n ") == ""
+    assert clean('""') == ""
+    print("ok  title cleaning")
+
+
+def test_title_request():
+    """Client.complete posts a non-streaming request and returns the text."""
+    seen = {}
+
+    class FakeResponse:
+        status_code = 200
+        text = ""
+
+        def raise_for_status(self):
+            pass
+
+        def json(self):
+            return {"choices": [{"message": {"content": " A Title "}}]}
+
+    def fake_post(url, json=None, timeout=None):
+        seen["url"] = url
+        seen["body"] = json
+        return FakeResponse()
+
+    import httpx
+
+    original = httpx.post
+    httpx.post = fake_post
+    try:
+        client = backend.Client("http://router:8181")
+        out = client.complete("m", [{"role": "user", "content": "hi"}], max_tokens=32)
+    finally:
+        httpx.post = original
+
+    assert out == " A Title ", out
+    assert seen["url"] == "http://router:8181/v1/chat/completions"
+    assert seen["body"]["model"] == "m"
+    assert seen["body"]["stream"] is False
+    assert seen["body"]["max_tokens"] == 32
+    # No tools are offered: a title turn must not trigger a search.
+    assert "tools" not in seen["body"]
+    # Thinking off, or a reasoning model spends the budget and returns "".
+    assert seen["body"]["chat_template_kwargs"] == {"enable_thinking": False}
+
+    # A reply with no choices is empty rather than an exception.
+    class Empty(FakeResponse):
+        def json(self):
+            return {}
+
+    httpx.post = lambda url, json=None, timeout=None: Empty()
+    try:
+        assert backend.Client("http://x").complete("m", []) == ""
+    finally:
+        httpx.post = original
+    print("ok  title request")
+
+
+def test_needs_title():
+    """Which finished turns are worth asking the model to name.
+
+    The trigger is a session still carrying its placeholder title, not the
+    turn number: a first reply that comes back empty (the follow-up
+    <tool_call> bug) must not forfeit titling for the whole session, and a
+    one-shot window keeps appending to the same session until New is
+    pressed.
+    """
+    needs = backend.needs_title
+    q = "how do I build a custom kernel on slackware?"
+
+    assert needs(q, q, "Fetch the sources.") is True
+    # Nothing to title from: the empty-reply turn is skipped, not consumed.
+    assert needs(q, q, "") is False
+    assert needs(q, q, "   \n ") is False
+    # Already named by the model, so leave it alone.
+    assert needs("Building a custom kernel", q, "Fetch the sources.") is False
+    # A long question is stored truncated; that still counts as untouched.
+    long_q = "x" * 200
+    assert needs(long_q[:60], long_q, "an answer") is True
+    # A title the user or model set that happens to be short is not a
+    # placeholder, even though it is under the cut-off.
+    assert needs("Kernels", long_q, "an answer") is False
+    print("ok  needs title")
+
+
+def test_title_prompt():
+    """The title request carries the exchange and asks for a short title."""
+    messages = backend.title_messages("how do I build a kernel?", "Run make.")
+    assert messages[-1]["role"] == "user"
+    blob = json.dumps(messages)
+    assert "how do I build a kernel?" in blob
+    assert "Run make." in blob
+    # A huge exchange is trimmed so titling never costs a full context.
+    big = backend.title_messages("x" * 10000, "y" * 10000)
+    assert len(json.dumps(big)) < 6000, len(json.dumps(big))
+    print("ok  title prompt")
+
+
 if __name__ == "__main__":
     test_presets()
     test_real_presets()
@@ -1504,4 +1650,9 @@ if __name__ == "__main__":
     test_search_storage()
     test_date_note()
     test_search_html()
+    test_final_round_note()
+    test_title_cleaning()
+    test_title_request()
+    test_needs_title()
+    test_title_prompt()
     print("\nall checks passed")