diff --git a/clearwing/sourcehunt/hunter.py b/clearwing/sourcehunt/hunter.py
index ef58527d..4dfbeaa9 100644
--- a/clearwing/sourcehunt/hunter.py
+++ b/clearwing/sourcehunt/hunter.py
@@ -1627,6 +1627,7 @@ async def arun(self) -> HunterRunResult:
         prev_step_had_skip = False
         while True:
             step += 1
+            synthesis_prompt_added = False
             progress_sig = (
                 len(self.ctx.findings),
                 len(self.ctx.potentials),
@@ -1648,6 +1649,7 @@ async def arun(self) -> HunterRunResult:
                 near_steps = step >= self.max_steps - 1
                 if near_budget or near_steps:
                     synthesis_injected = True
+                    synthesis_prompt_added = True
                     messages.append(
                         ChatMessage(
                             "user",
@@ -1663,6 +1665,7 @@ async def arun(self) -> HunterRunResult:
                     )
             final_synthesis_turn = step == self.max_steps
             if final_synthesis_turn:
+                synthesis_prompt_added = True
                 messages.append(
                     ChatMessage(
                         "user",
@@ -1671,7 +1674,17 @@ async def arun(self) -> HunterRunResult:
                         "recorded. If no finding was recorded, state that plainly.",
                     )
                 )
-            stop_reason = self._should_stop(step, total_cost_usd, steps_since_progress)
+            # A synthesis prompt must reach the model before the stall guard
+            # can terminate the hunt. Otherwise a late stall can enqueue the
+            # prompt and immediately exit without giving the model the final
+            # turn that was intended to conclude the investigation cleanly.
+            # Budget and max-step limits still win because only the progress
+            # counter is suppressed for this one turn.
+            stop_reason = self._should_stop(
+                step,
+                total_cost_usd,
+                0 if synthesis_prompt_added else steps_since_progress,
+            )
             if stop_reason:
                 logger.warning(
                     "Hunter stopped for %s: %s (step=%d, cost=$%.4f, findings=%d)",
diff --git a/tests/test_deep_agent_loop.py b/tests/test_deep_agent_loop.py
index 8def6586..4be5b5ae 100644
--- a/tests/test_deep_agent_loop.py
+++ b/tests/test_deep_agent_loop.py
@@ -995,3 +995,32 @@ async def test_hunter_stalls_when_no_progress():
 
     assert result.stop_reason == "stalled"
     assert llm.achat.call_count < 20  # stops well short of max_steps
+
+
+@pytest.mark.asyncio
+async def test_late_stall_delivers_final_synthesis_prompt_before_stopping():
+    hunter, llm = _make_hunter(agent_mode="deep", max_steps=4, budget_usd=0.0)
+    hunter.max_steps_without_progress = 3
+    counter = itertools.count()
+    llm.achat.side_effect = [
+        FakeResponse(
+            text="still investigating",
+            tool_calls_list=[_make_tool_call("think", {"notes": f"step {next(counter)}"})],
+        ),
+        FakeResponse(
+            text="still investigating",
+            tool_calls_list=[_make_tool_call("think", {"notes": f"step {next(counter)}"})],
+        ),
+        FakeResponse(text="Investigation complete; no finding was recorded."),
+    ]
+
+    with patch("clearwing.sourcehunt.hunter.HunterTrajectoryLogger") as mock_traj:
+        mock_traj.for_hunter.return_value = MagicMock()
+        result = await hunter.arun()
+
+    assert result.stop_reason == "completed"
+    assert llm.achat.call_count == 3
+    final_messages = llm.achat.call_args.kwargs["messages"]
+    assert any(
+        "approaching the end of your budget" in message.content for message in final_messages
+    )