feat: preserve planner context across task steps

tests/test_ai_planner.py

@@ -8,6 +8,7 @@ from core.errors import TaskFailedError
 from core.models import Bounds, Scene, SceneElement
 from runtime.ai_planner import AIPlanner
 from runtime.context import TaskContext
+from runtime.executor import StepResult
 from runtime.planner_config import PlannerConfig
 from runtime.tool_calling_client import ToolCallDecision
 from runtime.tool_specs import ALL_TOOL_SPECS
@@ -26,6 +27,7 @@ class FakeToolCallingClient:
         screenshot: bytes | None,
         tools: list[Any],
         timeout: float,
+        history: list[dict[str, Any]] | None = None,
     ) -> ToolCallDecision:
         self.calls.append(
             {
@@ -34,6 +36,7 @@ class FakeToolCallingClient:
                 "screenshot": screenshot,
                 "tools": tools,
                 "timeout": timeout,
+                "history": list(history) if history is not None else None,
             }
         )
         return self.decision
@@ -242,63 +245,64 @@ def test_ai_planner_propagates_rationale_and_thinking_to_planned_step() -> None:
     assert steps[0].expected_outcome == "The account settings page is visible."
 
 
-def test_history_summary_returns_compact_format() -> None:
-    from collections import deque
-    from runtime.ai_planner import _history_summary
-    from world.models import WorldEvent, WorldState
-
-    state = WorldState(
-        history=deque(
-            [
-                WorldEvent(
-                    action="tap",
-                    success=True,
-                    rationale="Opened settings.",
-                    arguments={"x": 1, "y": 2},
-                    purpose="Open settings.",
-                    expected_outcome="Settings is visible.",
-                    page="Home",
-                ),
-                WorldEvent(
-                    action="swipe",
-                    success=False,
-                    rationale=None,
-                    arguments={"start_y": 700, "end_y": 200},
-                    page="Settings",
-                ),
-            ]
+def test_ai_planner_carries_completed_turn_into_the_next_llm_call() -> None:
+    client = FakeToolCallingClient(
+        ToolCallDecision(
+            tool_name="tap",
+            arguments={"x": 1, "y": 2},
+            text_output="Opening the send control.",
+            purpose="Open the send control.",
+            expected_outcome="The composer is focused.",
         )
     )
+    planner = AIPlanner(client=client)
+    context = _context()
 
-    summary = _history_summary(state)
+    first_step = planner.plan(goal=context.goal, scene=_scene(), context=context)[0]
+    context.add_step_result(
+        StepResult(
+            step=first_step,
+            success=True,
+            attempts=1,
+            result={"ok": True},
+        )
+    )
+    planner.plan(goal=context.goal, scene=_scene(), context=context)
 
-    assert summary == [
+    assert client.calls[0]["history"] == []
+    history = client.calls[1]["history"]
+    assert history is not None
+    assert history[0]["tool_name"] == "tap"
+    assert history[0]["arguments"] == {
+        "x": 1,
+        "y": 2,
+        "purpose": "Open the send control.",
+        "expected_outcome": "The composer is focused.",
+    }
+    assert history[0]["tool_result"]["success"] is True
+    assert history[0]["tool_result"]["result"] == {"ok": True}
+
+
+def test_ai_planner_limits_history_sent_to_the_llm() -> None:
+    client = FakeToolCallingClient(
+        ToolCallDecision(tool_name="tap", arguments={"x": 1, "y": 2})
+    )
+    planner = AIPlanner(client=client, config=PlannerConfig(history_max_turns=2))
+    context = _context()
+    context.planner_history.extend(
         {
-            "page": "Home",
-            "action": "tap",
-            "arguments": {"x": 1, "y": 2},
-            "rationale": "Opened settings.",
-            "purpose": "Open settings.",
-            "expected_outcome": "Settings is visible.",
-            "success": True,
-        },
-        {
-            "page": "Settings",
-            "action": "swipe",
-            "arguments": {"start_y": 700, "end_y": 200},
+            "user_prompt": f"turn-{index}",
+            "tool_name": "tap",
+            "arguments": {},
             "rationale": None,
-            "purpose": None,
-            "expected_outcome": None,
-            "success": False,
-        },
+            "tool_result": {"success": True},
+        }
+        for index in range(3)
+    )
+
+    planner.plan(goal=context.goal, scene=_scene(), context=context)
+
+    assert [turn["user_prompt"] for turn in client.calls[0]["history"]] == [
+        "turn-1",
+        "turn-2",
     ]
-    # Must not contain scene element data
-    for entry in summary:
-        assert "scene_summary" not in entry
-        assert "elements" not in entry
-
-
-def test_history_summary_returns_empty_for_none_world() -> None:
-    from runtime.ai_planner import _history_summary
-
-    assert _history_summary(None) == []

tests/test_planner_prompts.py

@@ -9,7 +9,6 @@ def test_planner_user_prompt_includes_time_zone_and_configured_device_type() ->
     prompt = planner_user_prompt(
         goal="open settings",
         scene_json={"screen": {"width": 1, "height": 1}, "elements": []},
-        history_summary=[],
         device_platform="ios",
         now=datetime(
             2026,
@@ -35,8 +34,16 @@ def test_planner_user_prompt_uses_scene_platform_when_context_is_unavailable() -
             "elements": [],
             "app": {"platform": "android"},
         },
-        history_summary=[],
         now=datetime(2026, 7, 16, tzinfo=timezone.utc),
     )
 
     assert "Device type: android" in prompt
+
+
+def test_planner_user_prompt_does_not_duplicate_conversation_history() -> None:
+    prompt = planner_user_prompt(
+        goal="open settings",
+        scene_json={"screen": {"width": 1, "height": 1}, "elements": []},
+    )
+
+    assert "Recent history" not in prompt