fix: reject invalid and duplicate Work writes

tests/test_agent_loop.py

@@ -11,6 +11,7 @@ from agent_platform.runtime.loop import (
     AgentLoop,
     is_substantive_execution,
     normalize_write_file_content,
+    python_source_error,
     select_tool_specs,
     tool_event_details,
 )
@@ -39,6 +40,13 @@ def test_write_file_normalizes_double_escaped_plain_text() -> None:
     )
 
 
+def test_invalid_complete_python_source_is_detected_before_write() -> None:
+    error = python_source_error({"path": "benchmark.py", "content": "# Report\n- O(n²)\n"})
+    assert error is not None
+    assert "invalid character" in error
+    assert python_source_error({"path": "benchmark.py", "content": "print('ok')\n"}) is None
+
+
 def test_read_only_shell_commands_do_not_satisfy_execution_evidence() -> None:
     assert not is_substantive_execution("exec", {"command": "cat report.txt"})
     assert not is_substantive_execution("exec", {"command": "sed -n '1,20p' script.py"})
@@ -509,6 +517,49 @@ async def test_duplicate_commands_in_one_model_response_execute_once(settings) -
         await store.close()
 
 
+async def test_duplicate_identical_writes_execute_once(settings) -> None:
+    batch = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [
+                        tool_call("1", "write_file", {"path": "sort.py", "content": "print('ok')"}),
+                        tool_call("2", "write_file", {"path": "sort.py", "content": "print('ok')"}),
+                        tool_call("3", "exec", {"command": "python3 sort.py"}),
+                    ],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    final = {"choices": [{"message": {"role": "assistant", "content": "完成"}, "finish_reason": "stop"}]}
+    provider = ScriptedProvider([batch, final])
+    registry = RecordingRegistry()
+    store = RuntimeStore(settings.database_url)
+    await store.initialize()
+
+    async def callback(event_type, payload):
+        return None
+
+    try:
+        answer = await AgentLoop(provider, registry, store, max_tool_output_chars=10_000).run(
+            spec=get_model_spec("work-light"),
+            messages=[{"role": "user", "content": "写脚本并运行"}],
+            identity=UserIdentity("u1", "", "", "user"),
+            raw_user_jwt="jwt",
+            chat_id="c1",
+            callback=callback,
+        )
+        assert answer == "完成"
+        assert [name for name, _ in registry.calls] == ["write_file", "exec"]
+        duplicate = provider.requests[1]["messages"][-2]
+        assert duplicate["role"] == "tool"
+        assert "same path and identical content" in duplicate["content"]
+    finally:
+        await store.close()
+
+
 async def test_script_delivery_requires_successful_execution(settings) -> None:
     write = {
         "choices": [