feat: add official Codex skill and harden task execution

Install and update artemis-use with task diagnostics and plugin guidance. Preserve cross-device file receipts, review canonical reports, guard stage transitions and reduce small history summaries. Improve exploration lifecycle, request recovery and browser live connections.

tests/unit/agents/test_checker.py

@@ -673,7 +673,7 @@ def test_format_history_uses_session_offsets_and_skips_empty_steps():
     assert "'Login'" in lines[1]
 
 
-async def _final_check_human_text(ctx) -> str:
+async def _final_check_human_text(ctx, **kwargs) -> str:
     report = CheckReport(verdicts=[])
     structured = MagicMock()
     structured.ainvoke = AsyncMock(return_value=report)
@@ -691,7 +691,9 @@ async def _final_check_human_text(ctx) -> str:
             new=AsyncMock(return_value=(None, "elements")),
         ),
     ):
-        await run_final_check(ctx, goal="the goal", plan_text="- [x] G", ledger=[], check_items=[])
+        await run_final_check(
+            ctx, goal="the goal", plan_text="- [x] G", ledger=[], check_items=[], **kwargs
+        )
 
     messages = acomplete_mock.await_args.args[1]
     human = messages[-1]
@@ -730,6 +732,56 @@ async def test_final_review_reports_no_guidance_explicitly():
     assert "# User guidance received during the run\n(none)\n" in text
 
 
+@pytest.mark.asyncio
+async def test_final_review_receives_verified_file_receipts_without_history_tool_call():
+    ctx = _mock_ctx()
+    ctx.data_engine.get_agent_friendly_steps.return_value = [
+        {
+            "step_number": 44,
+            "summary": "已写入报告",
+            "action_taken": {"action": "android_file"},
+            "last_execution_result": {
+                "result": '{"ok":true,"verified":true,"device_path":"/sdcard/report.md","bytes":8022}'
+            },
+        }
+    ]
+    text = await _final_check_human_text(ctx, execution_report="报告保存情况未知")
+    assert "Recorded File Receipts (tool evidence, not report claims)" in text
+    assert '"step": 44' in text and '"verified": true' in text
+    assert "/sdcard/report.md" in text
+
+
+@pytest.mark.parametrize("entry", ["checkpoint", "final"])
+def test_visual_acceptance_requires_counter_evidence_without_exempting_real_defects(entry):
+    text = assemble_checker_prompt_segments(entry, [], probe_tool_registered=False)
+    assert "compare recorded pre/post screenshots and nearby steps" in text
+    assert "user permits this behavior" in text
+    assert "fixed navigation is not a blanket exemption" in text
+    assert "explicit visual standard" in text
+    assert "inconclusive, not failed or automatically passed" in text
+
+
+@pytest.mark.asyncio
+async def test_final_review_distinguishes_actual_delivery_from_report_claims():
+    text = await _final_check_human_text(
+        _mock_ctx(),
+        execution_report="# 完整报告\n实际发现与未测范围",
+        report_delivery={"report_path": "notes/output.md", "exploration_message_sequences": [12]},
+    )
+    assert "# 完整报告" in text
+    assert "Report Delivery Receipt (recorded by ARTEMIS, not a model claim)" in text
+    assert '"exploration_message_sequences": [12]' in text
+    assert "confirms delivery only" in text
+    assert "against actual device/tool evidence" in text
+    without_receipt = await _final_check_human_text(_mock_ctx(), execution_report="模型自述已交付")
+    assert "Report Delivery Receipt" not in without_receipt
+    failed = await _final_check_human_text(
+        _mock_ctx(), report_delivery={"ok": False, "verified": False, "error": "disk full"}
+    )
+    assert "whether the executor's report was saved" in failed
+    assert '"verified": false' in failed and "disk full" in failed
+
+
 @pytest.mark.asyncio
 async def test_exploration_review_sees_conversation_and_can_read_earlier_messages():
     ctx = _mock_ctx()