An empty plan against an unmet goal gets one nudge, then is believed

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: Shashankss1205

grapharc/planner/loop.py

@@ -392,6 +392,10 @@ def run(
         unplanned_in_a_row = 0
         failed_in_a_row = 0
         stalled_in_a_row = 0
+        # One nudge per run when an empty proposal contradicts an unmet goal
+        # check; `confirm_pending` marks the single round right after it.
+        empty_nudged = False
+        confirm_pending = False
 
         stop, detail = self._precheck(current)
         round_number = 0
@@ -441,6 +445,19 @@ def close(**fields: Any) -> None:
                 break
 
             if not outcome.ok or outcome.proposal is None:
+                if confirm_pending:
+                    # The planner already said "nothing more"; the nudge asked
+                    # it to confirm and the follow-up produced nothing usable.
+                    # Read that as the confirmation it is, not as a planning
+                    # failure to burn the allowance on — a scripted planner
+                    # whose replies simply ran out lands here too.
+                    stop = LoopStop.NO_FURTHER_WORK
+                    detail = (
+                        "the planner proposed no further work, and the "
+                        "follow-up produced nothing usable"
+                    )
+                    close(planner_error=outcome.error, tokens=outcome.tokens)
+                    break
                 unplanned_in_a_row += 1
                 # The retry note shows the model what it actually said and what
                 # was wanted. The bare error string alone ("no JSON object
@@ -466,6 +483,7 @@ def close(**fields: Any) -> None:
                 close(planner_error=outcome.error, tokens=outcome.tokens)
                 continue
             unplanned_in_a_row = 0
+            confirm_pending = False
 
             proposal = outcome.proposal
             verdict = self.checker.check(
@@ -490,11 +508,36 @@ def close(**fields: Any) -> None:
 
             if not proposal.nodes:
                 # Admitted, and it authorises nothing: the planner is saying
-                # there is no further work. Admitting that is the right answer.
+                # there is no further work. With no goal check, or a satisfied
+                # one, admitting that is the right answer. With an UNMET goal
+                # check it contradicts the operator's own definition of done,
+                # so it gets one nudge naming the contradiction — one, not a
+                # counter, because a planner that says "nothing more" twice is
+                # answering, not failing.
+                if (
+                    self.goal_reached is not None
+                    and not self._goal_met(current)
+                    and not empty_nudged
+                ):
+                    empty_nudged = True
+                    confirm_pending = True
+                    note = (
+                        "Your previous proposal was empty, but the run's goal "
+                        "check is not yet satisfied. Propose the remaining "
+                        "work, or reply with an empty proposal again to "
+                        "confirm there is nothing more this catalog can do."
+                    )
+                    close(**judged)
+                    continue
                 stop = (
                     LoopStop.GOAL_MET if self._goal_met(current) else LoopStop.NO_FURTHER_WORK
                 )
-                detail = "the planner proposed no further work"
+                detail = (
+                    "the planner confirmed no further work; the goal check is "
+                    "still unsatisfied"
+                    if empty_nudged and not self._goal_met(current)
+                    else "the planner proposed no further work"
+                )
                 close(**judged)
                 break
 

tests/test_planner_loop.py

@@ -1661,3 +1661,72 @@ def test_the_incident_example_state_merges_parallel_writers():
     result = loop.run("triage, patch and verify at once", IncidentState())
     assert result.stop.value == "goal_met"
     assert sorted(result.state.notes) == ["patch ran", "triage ran", "verify ran"]
+
+
+def test_an_empty_proposal_against_an_unmet_goal_gets_one_nudge():
+    """Empty plan, goal check unsatisfied: the contradiction is named once.
+
+    The planner is told the goal check is not met and asked to either propose
+    the remaining work or repeat the empty proposal. Here it proposes the
+    work, and the run finishes on the goal — where before the nudge existed,
+    round 1's empty reply ended the run as `no_further_work` with the goal
+    never mentioned to the model.
+    """
+    loop, model, bodies = build_loop(
+        [NOTHING_MORE, plan(("write", "summarise"))], goal_reached=goal_is_done
+    )
+
+    result = loop.run("summarise the findings", LoopState())
+
+    assert result.stop is LoopStop.GOAL_MET
+    assert bodies.ran == ["write"]
+    assert len(result.rounds) == 2
+    assert not result.rounds[0].executed and result.rounds[0].admitted
+    assert any(
+        "goal check is not yet satisfied" in str(message.content)
+        for message in model.calls[1]
+    )
+
+
+def test_a_second_empty_proposal_is_believed():
+    """The nudge is one round, not a counter: a repeat empty plan is an answer."""
+    loop, model, bodies = build_loop(
+        [NOTHING_MORE, NOTHING_MORE], goal_reached=goal_is_done
+    )
+
+    result = loop.run("summarise the findings", LoopState())
+
+    assert result.stop is LoopStop.NO_FURTHER_WORK
+    assert "confirmed no further work" in result.detail
+    assert bodies.ran == []
+    assert len(result.rounds) == 2
+    assert model.call_count == 2
+
+
+def test_an_unusable_reply_after_the_nudge_confirms_no_further_work():
+    """A planner with nothing left to say after the nudge is not a failure.
+
+    The scripted stand-ins end their reply lists with an empty proposal; the
+    nudge asks one more question than the script answers. Exhaustion there
+    must read as the confirmation it is — never as `planning_failed` burning
+    the failure allowance on a planner that already said "nothing more".
+    """
+    loop, model, bodies = build_loop([NOTHING_MORE], goal_reached=goal_is_done)
+
+    result = loop.run("summarise the findings", LoopState())
+
+    assert result.stop is LoopStop.NO_FURTHER_WORK
+    assert "nothing usable" in result.detail
+    assert bodies.ran == []
+    assert result.rounds[-1].planner_error
+
+
+def test_an_empty_proposal_with_no_goal_check_stops_without_a_nudge():
+    """No goal check means nothing to contradict: one round, one clean stop."""
+    loop, model, bodies = build_loop([NOTHING_MORE])
+
+    result = loop.run("nothing needs doing", LoopState())
+
+    assert result.stop is LoopStop.NO_FURTHER_WORK
+    assert len(result.rounds) == 1
+    assert model.call_count == 1