A ceiling for a chat, and a nudge for an agent that stops early

MAX_ROUNDS = 1 was wrong, and wrong in a way worth writing down. The loop
already ends the moment a round comes back with no tool calls -- that is the
model saying it has what it needs, and it is the termination condition every
agentic harness uses. A round limit was never a schedule; it exists to catch the
case where the model never says so. One is low enough to stop being a ceiling
and start being a schedule: it overrode the model's judgement on every single
turn.

And it broke something concrete. Several built-ins are two-step pairs --
knowledge_get and notes_get read a document "by the id a search returned" -- so
one round left the library searchable and not readable. That is not an edge
case, it is the library working at half depth, and I understated it as "cannot
search the web and then read a result" when the change went in.

It is a setting now, under General, default 5, with 0 meaning no ceiling. The
loop and the harness both read settings_store.chat_rounds, so the model is never
told a budget that is not its own; tools.MAX_ROUNDS is the fallback for callers
with no session and a test pins the two equal. core.rounds goes back to naming
the number, and vanishes entirely when there is no ceiling rather than promising
zero rounds.

The other half of "let it decide how long to go": an agent reply that ends while
its plan still has open tasks is asked once to carry on. Only against a plan,
because that is the one thing there is to be objectively wrong about -- a model
with no plan that says it has finished is believed, and arguing with it would be
guessing. At most twice in a row, with the count reset the moment it calls a
tool again, so the bound is on consecutive stops rather than on stops in total.
Never in Plan mode and never past plan_submit, which ends the turn on purpose.
Giving up is recorded as an event rather than left silent.

The model's own words go back with the nudge, which turned up a real bug on the
way: ReasoningSplitter holds back a few characters against a <think> tag split
across chunks, so round_text at the end of a round was missing its tail. That
text is echoed as an assistant turn for tool rounds too, so a model has been
occasionally asked to continue from a transcript where it trailed off
mid-sentence. Flushed per round now.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

tests/test_agent_plan.py

@@ -351,3 +351,160 @@ def test_a_plan_pointer_at_another_chats_message_is_ignored(db, owner):
     db.commit()
 
     assert session.resolve(db, chat, owner).plan == {}
+
+
+# --- Being asked to carry on ------------------------------------------------------
+def _stub(rounds, seen):
+    async def stream_chat(_endpoint, payload):
+        seen.append(payload)
+        for chunk in rounds[min(len(seen) - 1, len(rounds) - 1)]:
+            yield chunk
+
+    return stream_chat
+
+
+def _text(text):
+    return {"choices": [{"delta": {"content": text}}]}
+
+
+def _call(name, arguments):
+    return {"choices": [{"delta": {"tool_calls": [
+        {"index": 0, "id": "c1", "function": {"name": name, "arguments": arguments}}]}}]}
+
+
+def _reply(db, chat):
+    from lembas.services import generation as generation_service
+
+    db.add(Message(chat_id=chat.id, role="user", content="do it", complete=True))
+    db.commit()
+    assistant = Message(chat_id=chat.id, role=ROLE_ASSISTANT, content="", complete=False)
+    db.add(assistant)
+    db.commit()
+    return generation_service.Generation(chat_id=chat.id, message_id=assistant.id)
+
+
+async def _run(monkeypatch, generation, rounds, seen):
+    from lembas.services import generation as generation_service
+
+    monkeypatch.setattr(generation_service, "stream_chat", _stub(rounds, seen))
+
+    async def _no_title(*_a, **_k):
+        return ""
+
+    monkeypatch.setattr("lembas.services.chat.generate_title", _no_title)
+    await generation_service._run(generation)
+
+
+async def test_stopping_with_open_tasks_is_answered_with_carry_on(db, owner, monkeypatch):
+    """The half prompting cannot do. core.keep_working tells it not to stop
+    halfway; this is what happens when it does anyway."""
+    chat = _agent_chat(db, owner, mode=policy.MODE_AUTO)
+    _with_plan(db, chat, V1)
+    generation = _reply(db, chat)
+
+    seen: list[dict] = []
+    await _run(monkeypatch, generation, [[_text("I have done the first bit.")],
+                                         [_text("All finished.")]], seen)
+
+    # Three requests: the reply, then a nudge, then a second nudge -- this
+    # model never marks anything done, so the plan stays open and it is asked
+    # until MAX_NUDGES runs out. That it gives up is the next test.
+    assert len(seen) == 3, "it was asked again"
+    nudge = seen[1]["messages"][-1]
+    assert nudge["role"] == "user"
+    assert "still has work in it" in nudge["content"]
+    assert "Rotate it" in nudge["content"], "and says which tasks"
+    # Its own words go back with it, or it is asked to carry on from a
+    # transcript in which it never spoke.
+    assert seen[1]["messages"][-2]["content"] == "I have done the first bit."
+
+
+async def test_a_finished_plan_is_believed(db, owner, monkeypatch):
+    chat = _agent_chat(db, owner, mode=policy.MODE_AUTO)
+    done = plans.build(title="x", phases=[{"title": "P", "tasks": ["a"]}])
+    done["phases"][0]["tasks"][0]["status"] = "done"
+    _with_plan(db, chat, done)
+    generation = _reply(db, chat)
+
+    seen: list[dict] = []
+    await _run(monkeypatch, generation, [[_text("All done.")]], seen)
+
+    assert len(seen) == 1
+
+
+async def test_a_chat_with_no_plan_is_never_nudged(db, owner, monkeypatch):
+    """There is nothing to be objectively wrong about, so a model that says it
+    has finished is believed."""
+    chat = _agent_chat(db, owner, mode=policy.MODE_AUTO)
+    generation = _reply(db, chat)
+
+    seen: list[dict] = []
+    await _run(monkeypatch, generation, [[_text("All done.")]], seen)
+
+    assert len(seen) == 1
+
+
+async def test_plan_mode_is_never_nudged(db, owner, monkeypatch):
+    """plan_submit ends the turn deliberately. Nudging past it would argue with
+    the whole point of the mode."""
+    chat = _agent_chat(db, owner, mode=policy.MODE_PLAN)
+    _with_plan(db, chat, V1)
+    generation = _reply(db, chat)
+
+    seen: list[dict] = []
+    await _run(monkeypatch, generation, [[_text("Here is what I would do.")]], seen)
+
+    assert len(seen) == 1
+
+
+async def test_it_gives_up_after_two_and_says_so(db, owner, monkeypatch):
+    """A model that has nothing left to do must be able to say so and be
+    believed rather than argued with indefinitely -- and a reply that stopped
+    twice with work outstanding is worth being able to see afterwards."""
+    chat = _agent_chat(db, owner, mode=policy.MODE_AUTO)
+    _with_plan(db, chat, V1)
+    generation = _reply(db, chat)
+
+    seen: list[dict] = []
+    await _run(monkeypatch, generation, [[_text("Nothing more from me.")]], seen)
+
+    from lembas.services import generation as generation_service
+
+    assert len(seen) == generation_service.MAX_NUDGES + 1
+    assert generation.tool_events[-1]["status"] == "error"
+    assert "asked twice" in generation.tool_events[-1]["error"]
+
+
+async def test_calling_a_tool_again_resets_the_count(db, owner, monkeypatch):
+    """The count is of consecutive stops. A model that stops, is nudged, does
+    some work and stops again has not run out of patience."""
+    chat = _agent_chat(db, owner, mode=policy.MODE_AUTO)
+    _with_plan(db, chat, V1)
+    generation = _reply(db, chat)
+
+    seen: list[dict] = []
+    await _run(
+        monkeypatch,
+        generation,
+        [[_text("Pausing.")], [_call("file_list", '{"path": "."}')], [_text("Pausing again.")]],
+        seen,
+    )
+
+    # Without the reset this run would end after two nudges, at three requests.
+    # The tool call in the middle clears the count, so it gets more than that --
+    # which is the property, and does not depend on where the stub stops.
+    assert len(seen) > 3
+
+
+async def test_the_switch_turns_it_off(db, owner, monkeypatch):
+    from lembas.services import settings_store
+
+    settings_store.update(db, {"nudge_unfinished": False}, key=settings_store.AGENTS)
+    chat = _agent_chat(db, owner, mode=policy.MODE_AUTO)
+    _with_plan(db, chat, V1)
+    generation = _reply(db, chat)
+
+    seen: list[dict] = []
+    await _run(monkeypatch, generation, [[_text("All done.")]], seen)
+
+    assert len(seen) == 1