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/graph/test_exit_settlement.py

@@ -17,6 +17,7 @@ settlement timeout, when-semantics at the exit, run outcome, and the four
 switch-combination matrix."""
 
 import asyncio
+import json
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
@@ -83,7 +84,7 @@ async def test_exploration_question_prevents_completed_plan_from_starting_final_
     repo.ask(run["id"], "下一步做什么？", "返回主页", "ask")
     with patch("artemis.graph.graph.run_final_check", new=AsyncMock()) as reviewer:
         assert convergence_gate(_make_state(), ctx) == "continue"
-        result = await exit_settlement_node(_make_state(), ctx)
+        result = await exit_settlement_node(_make_state(execution_report="本次调查报告"), ctx)
     assert result == {"exit_settlement_route": "continue"}
     reviewer.assert_not_called()
     assert repo.get(session["id"])["pending_question"]
@@ -100,7 +101,7 @@ async def test_exploration_late_return_home_guidance_invalidates_final_review(ex
         return CheckReport(verdicts=[], unmet_subgoals=[])
 
     with patch("artemis.graph.graph.run_final_check", new=review):
-        result = await exit_settlement_node(_make_state(), ctx)
+        result = await exit_settlement_node(_make_state(execution_report="本次调查报告"), ctx)
     assert result == {"exit_settlement_route": "continue"}
     assert repo.run(run["id"])["status"] == "running"
     assert repo.run(run["id"])["user_guidance_count"] == 1
@@ -118,12 +119,99 @@ async def test_review_stops_capture_before_waiting_and_seals_exploration(explora
         return CheckReport(verdicts=[], unmet_subgoals=[])
 
     with patch("artemis.graph.graph.run_final_check", new=review):
-        result = await exit_settlement_node(_make_state(), ctx)
+        result = await exit_settlement_node(_make_state(execution_report="本次调查报告"), ctx)
     assert result["exit_settlement_route"] == "end"
     assert repo.get(session["id"])["status"] == "finalizing"
     ctx.on_execution_resume.assert_not_called()
 
 
+@pytest.mark.asyncio
+async def test_completed_exploration_waits_for_report_before_review(exploration_exit):
+    ctx, repo, session, run = exploration_exit
+    state = State.initial("保存并交付完整报告")
+    with patch("artemis.graph.graph.run_final_check", new=AsyncMock()) as review:
+        assert convergence_gate(state, ctx) == "continue"
+        assert await exit_settlement_node(state, ctx) == {"exit_settlement_route": "continue"}
+    review.assert_not_called()
+    ctx.on_execution_review.assert_not_called()
+    assert repo.run(run["id"])["status"] == "running"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("verdict_status", ["passed", "failed", "inconclusive"])
+async def test_report_is_saved_and_fully_delivered_before_independent_review(
+    exploration_exit, verdict_status
+):
+    from artemis.runtime.case_result import CaseTestResult
+
+    ctx, repo, session, run = exploration_exit
+    body = "# 本次调查报告\n" + "实际证据及未测范围。" * 2400
+    state = State(initial_goal="交付完整报告", execution_report=body)
+    _write_plan(ctx.data_engine.base_dir, "- [x] 调查并交付报告\n  - assert: 实际页面结果\n")
+
+    async def review(*args, **kwargs):
+        directory = ctx.data_engine.base_dir
+        assert (directory / "notes" / "output.md").read_text(encoding="utf-8") == body
+        assert json.loads((directory / "result.json").read_text(encoding="utf-8")) == {
+            "result": body
+        }
+        delivered = [m for m in repo.messages(session["id"]) if m.get("kind") == "execution_report"]
+        assert len(delivered) == 3
+        assert "".join(m["text"] for m in delivered) == body
+        assert kwargs["execution_report"] == body
+        assert kwargs["report_delivery"]["exploration_message_sequences"] == [
+            m["sequence"] for m in delivered
+        ]
+        return CheckReport(
+            verdicts=[
+                CheckVerdict(
+                    item_text="实际页面结果",
+                    kind="assert",
+                    status=verdict_status,
+                    evidence="步骤 1",
+                )
+            ],
+            test_result=CaseTestResult(
+                status=verdict_status, reason="按本次实际证据裁定", evidence=["步骤 1 实际页面"]
+            ),
+        )
+
+    with patch("artemis.graph.graph.run_final_check", new=review):
+        result = await exit_settlement_node(state, ctx)
+    assert result["run_outcome"]["test_result"]["status"] == verdict_status
+    assert repo.run(run["id"])["status"] == "finalizing"
+
+
+@pytest.mark.asyncio
+async def test_report_delivery_is_idempotent_with_review_disabled(exploration_exit):
+    ctx, repo, session, run = exploration_exit
+    ctx.execution_setup.disable_checker = True
+    state = State(initial_goal="调查", execution_report="完整报告正文")
+    assert convergence_gate(state, ctx) == "exit_settlement"
+    with patch("artemis.graph.graph.run_final_check", new=AsyncMock()) as review:
+        await exit_settlement_node(state, ctx)
+        await exit_settlement_node(state, ctx)
+    review.assert_not_called()
+    delivered = [m for m in repo.messages(session["id"]) if m.get("kind") == "execution_report"]
+    assert len(delivered) == 1
+    assert delivered[0]["text"] == "完整报告正文"
+
+
+@pytest.mark.asyncio
+async def test_report_write_failure_cannot_release_as_completed(exploration_exit, monkeypatch):
+    from pathlib import Path
+
+    ctx, repo, session, run = exploration_exit
+    state = State(initial_goal="调查", execution_report="完整报告正文")
+    monkeypatch.setattr(Path, "replace", MagicMock(side_effect=OSError("disk full")))
+    with patch("artemis.graph.graph.run_final_check", new=AsyncMock()) as review:
+        with pytest.raises(OSError, match="disk full"):
+            await exit_settlement_node(state, ctx)
+    review.assert_not_called()
+    assert repo.run(run["id"])["status"] == "running"
+    assert not list(ctx.data_engine.base_dir.rglob("*.tmp"))
+
+
 def _make_ctx(tmp_path, **setup_kwargs):
     setup_kwargs.setdefault("disable_checker", False)
     ctx = MagicMock(spec=ArtemisContext)
@@ -311,7 +399,16 @@ async def test_final_check_receives_ledger_and_at_end_items(tmp_path):
 
     captured = {}
 
-    async def fake_final(ctx_arg, goal, plan_text, ledger, check_items, attempt_id=None):
+    async def fake_final(
+        ctx_arg,
+        goal,
+        plan_text,
+        ledger,
+        check_items,
+        attempt_id=None,
+        execution_report=None,
+        report_delivery=None,
+    ):
         captured["goal"] = goal
         captured["ledger"] = ledger
         captured["items"] = list(check_items)