update supervisor nudging logic and fix tests

src/lilith_agent/app.py

@@ -80,6 +80,7 @@ _FAIL_SAFE_RECURSION_HEADROOM = 4
 _SUPERVISOR_MIN_TOOL_CALLS = 5
 _SUPERVISOR_RECENT_MESSAGES = 12
 _SUPERVISOR_REVIEW_MAX = 3
+_SUPERVISOR_MAX_NUDGES = 5
 
 
 _RESPONSE_METADATA_NOISE_KEYS = frozenset({
@@ -828,6 +829,12 @@ def build_react_agent(cfg: Config):
         return {"messages": [response]}
 
     def supervisor_node(state):
+        if state.get("supervisor_nudges", 0) >= _SUPERVISOR_MAX_NUDGES:
+            return {
+                "supervisor_decision": "finalize",
+                "supervisor_best_answer": state.get("supervisor_best_answer", ""),
+                "supervisor_guidance": "Nudge cap reached. Commit to best available answer.",
+            }
         tool_calls_this_turn = _count_tool_calls_since_last_human(state["messages"])
         if supervisor_model is None or tool_calls_this_turn < _SUPERVISOR_MIN_TOOL_CALLS:
             return {
@@ -847,8 +854,9 @@ def build_react_agent(cfg: Config):
             "Return ONLY JSON with keys status, best_answer, guidance. "
             "status must be continue, nudge, or finalize. Use nudge when evidence likely supports "
             "an answer but one more agent turn is acceptable. Use finalize only when a concrete "
-            "submit-ready answer is available or the evidence is conclusive. A prior nudge alone is "
-            "not a reason to finalize. Never put placeholder values like unknown, n/a, none, or not "
+            "submit-ready answer is available or the evidence is conclusive. If you have already nudged "
+            "five times and the agent has not improved, use finalize to force termination — do not nudge "
+            "repeatedly. Never put placeholder values like unknown, n/a, none, or not "
             "sure in best_answer. If no concrete submit-ready answer is available, set best_answer "
             "to an empty string and use guidance to force the agent to make its best guess."
         )
@@ -863,8 +871,6 @@ def build_react_agent(cfg: Config):
         if _is_placeholder_answer(best_answer):
             print(f"[supervisor] discarded placeholder best_answer={best_answer[:80]!r}", flush=True)
             best_answer = ""
-        if status == "finalize" and state.get("supervisor_nudges", 0) > 0 and not best_answer:
-            status = "nudge"
         log.info(
             "[supervisor] status=%s best=%r guidance=%r",
             status,

tests/test_graph.py

@@ -428,15 +428,13 @@ def test_supervisor_finalizer_rejects_unknown_best_answer_and_forces_best_guess(
     assert strong.finalizer_calls == 1
 
 
-def test_supervisor_does_not_finalize_placeholder_after_prior_nudge(monkeypatch, tmp_path):
+def test_supervisor_finalizes_even_with_placeholder_if_requested(monkeypatch, tmp_path):
     class FakeBoundModel:
         def __init__(self):
             self.calls = 0
 
         def invoke(self, messages):
             self.calls += 1
-            if self.calls >= 3 and any("SUPERVISOR" in str(getattr(m, "content", "")) for m in messages):
-                return AIMessage(content="Final Answer: recovered guess")
             return _ai_with_calls([
                 {
                     "id": f"call-{self.calls}",
@@ -458,10 +456,11 @@ def test_supervisor_does_not_finalize_placeholder_after_prior_nudge(monkeypatch,
             prompt = str(messages[0].content)
             if "SUPERVISOR FINALIZER" in prompt:
                 self.finalizer_calls += 1
-                return AIMessage(content="Final Answer: premature finalizer")
+                return AIMessage(content="Final Answer: finalizer output")
             self.supervisor_calls += 1
             if self.supervisor_calls == 1:
                 return AIMessage(content='{"status":"nudge","best_answer":"","guidance":"Use the evidence to make a best guess."}')
+            # On second call, it asks to finalize but with a placeholder answer. The code should allow finalization.
             return AIMessage(content='{"status":"finalize","best_answer":"Unknown","guidance":"You were already nudged. Provide your final answer based on the best available information."}')
 
     strong = FakeStrongModel()
@@ -480,15 +479,15 @@ def test_supervisor_does_not_finalize_placeholder_after_prior_nudge(monkeypatch,
     graph = build_react_agent(cfg)
     result = graph.invoke(
         {"messages": [HumanMessage(content="Question requiring a concrete answer")], "iterations": 0, "todos": []},
-        {"configurable": {"thread_id": "supervisor-placeholder-after-nudge-test"}},
+        {"configurable": {"thread_id": "supervisor-finalize-after-nudge-test"}},
     )
 
-    assert result["messages"][-1].content == "Final Answer: recovered guess"
-    assert strong.finalizer_calls == 0
+    assert result["messages"][-1].content == "Final Answer: finalizer output"
+    assert strong.finalizer_calls == 1
     assert strong.supervisor_calls == 2
 
 
-def test_supervisor_does_not_auto_finalize_concrete_best_answer_after_prior_nudge(monkeypatch, tmp_path):
+def test_supervisor_forces_finalize_after_max_nudges(monkeypatch, tmp_path):
     class FakeBoundModel:
         def __init__(self):
             self.calls = 0
@@ -516,18 +515,19 @@ def test_supervisor_does_not_auto_finalize_concrete_best_answer_after_prior_nudg
             prompt = str(messages[0].content)
             if "SUPERVISOR FINALIZER" in prompt:
                 self.finalizer_calls += 1
-                return AIMessage(content="Final Answer: should not happen")
+                return AIMessage(content="Final Answer: max nudges forced this")
             self.supervisor_calls += 1
             return AIMessage(content='{"status":"nudge","best_answer":"concrete candidate","guidance":"Check one more constraint before final answer."}')
 
     strong = FakeStrongModel()
     cfg = Config.from_env()
-    cfg.recursion_limit = 10
+    cfg.recursion_limit = 20
     cfg.budget_hard_cap = 99
     cfg.budget_warn_at = 99
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 1, raising=False)
+    monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MAX_NUDGES", 5, raising=False)
     monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: object())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
@@ -536,12 +536,13 @@ def test_supervisor_does_not_auto_finalize_concrete_best_answer_after_prior_nudg
     graph = build_react_agent(cfg)
     result = graph.invoke(
         {"messages": [HumanMessage(content="Question requiring more checking")], "iterations": 0, "todos": []},
-        {"configurable": {"thread_id": "supervisor-no-auto-finalize-after-nudge-test"}},
+        {"configurable": {"thread_id": "supervisor-max-nudges-test"}},
     )
 
     assert strong.finalizer_calls == 0
-    assert strong.supervisor_calls > 1
-    assert result["supervisor_decision"] == "nudge"
+    assert result["messages"][-1].content == "Final Answer: concrete candidate"
+    assert strong.supervisor_calls == 5
+    assert result["supervisor_decision"] == "finalize"
 
 
 def test_final_answer_gets_supervisor_review_and_can_be_returned_for_revision(monkeypatch, tmp_path):