feat: 废弃 FC，改用纯 JSON action 模式调用工具

- engine.generate_response 不再下发 tools schema（TUI/CLI 分支均 tools=None） - 系统提示改为纯 JSON 调用格式说明，保留 tool_search 自查用法 - 新增 _loads_json 容错解析，容忍模型写出的 Windows 反斜杠路径（如 C:/Users） - 新增 tests/test_pure_json.py 覆盖解析容错/引擎无 schema 下发/端到端按名分发 Co-Authored-By: WorkBuddy <workbuddy@tencent.com>
2026-08-06 20:43:39 +08:00
1 parent b577b0ff21
commit c87a07706e
@@ -0,0 +1,145 @@
+"""
+纯 JSON 工具调用模式 测试
+
+取代 FC：引擎不再收到 tools schema，模型按系统提示里的
+{"action":"use_tool","tool":名,"args":参数} 文本输出，由 _parse_mixed_response
+解析、按名分发。重点验证：
+1. _loads_json / _parse_mixed_response 对 Windows 反斜杠路径健壮
+2. 引擎在 tools=None 时返回纯文本（不产 _fc / tool_calls）
+3. 端到端：真实 AICLI + FakeEngine，use_tool 被按名分发执行，且无 FC schema 下发
+"""
+import os
+import sys
+import json
+import tempfile
+
+import pytest
+
+PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+if PROJECT_ROOT not in sys.path:
+    sys.path.insert(0, PROJECT_ROOT)
+
+
+# ════════════════════════════════════════════════
+#  1. 解析器：Windows 反斜杠路径容错
+# ════════════════════════════════════════════════
+
+@pytest.fixture
+def cli():
+    from xcli_core.cli_core import AICLI
+    return AICLI()
+
+
+class TestParseRobustness:
+    def test_loads_json_tolerates_backslash_path(self, cli):
+        # 模型常把 Windows 路径写成单反斜杠 → 非法 JSON 转义
+        raw = r'{"action": "use_tool", "tool": "code_editor", "args": "create C:\Users\me\demo.py print(1)"}'
+        data = cli._loads_json(raw)
+        assert data["action"] == "use_tool"
+        assert data["tool"] == "code_editor"
+        # 修复后应是合法的单反斜杠路径
+        assert data["args"] == r"create C:\Users\me\demo.py print(1)"
+
+    def test_parse_mixed_extracts_use_tool_with_backslash(self, cli):
+        resp = (
+            '好的，我来创建文件。\n'
+            r'{"action": "use_tool", "tool": "code_editor", "args": "create C:\Users\me\demo.py print(1)"}'
+        )
+        text, data = cli._parse_mixed_response(resp)
+        assert isinstance(data, dict)
+        assert data["action"] == "use_tool"
+        assert data["tool"] == "code_editor"
+        assert r"C:\Users\me\demo.py" in data["args"]
+
+    def test_parse_mixed_plain_text_no_action(self, cli):
+        text, data = cli._parse_mixed_response("今天天气不错")
+        assert data is None
+        assert "今天天气不错" in text
+
+
+# ════════════════════════════════════════════════
+#  2. 引擎：tools=None 返回纯文本，不产 tool_calls
+# ════════════════════════════════════════════════
+
+class TestEngineToolsNone:
+    def test_openai_engine_no_tools_sent(self, monkeypatch):
+        import types
+        import ai_engines.openai_engine as oe
+
+        calls = []
+
+        def fake_post(url=None, json=None, headers=None, timeout=None, **kw):
+            calls.append(dict(json))
+            return types.SimpleNamespace(
+                status_code=200,
+                json=lambda: {"choices": [{"message": {"content": "好"}}],
+                               "usage": {"prompt_tokens": 5}},
+                text="{}",
+            )
+
+        monkeypatch.setattr(oe, "requests",
+                             types.SimpleNamespace(post=fake_post, exceptions=oe.requests.exceptions))
+
+        eng = oe.OpenaiAI.__new__(oe.OpenaiAI)
+        eng.name = "openai"
+        eng.api_key = "k"
+        eng.base_url = "https://example.com/v1"
+        eng.model = "gpt-4o"
+        eng.max_history = 10
+        eng.conversation_history = []
+        eng.shared_conversation_history = None
+        eng.last_prompt_tokens = None
+        eng.thinking_start_marker = "<think>"
+        eng.thinking_end_marker = "</think>"
+
+        reply = eng.generate_response("hi", tools=None)
+        assert reply == "好"
+        assert "tools" not in calls[0], "tools=None 时不应下发 FC schema"
+        assert "tool_choice" not in calls[0]
+
+
+# ════════════════════════════════════════════════
+#  3. 端到端：真实 AICLI + FakeEngine，use_tool 按名分发
+# ════════════════════════════════════════════════
+
+class FakeEngine:
+    def __init__(self):
+        self.calls = 0
+        self.tools_seen = []
+
+    @property
+    def name(self):
+        return "fake"
+
+    def generate_response(self, user_input, tool_results=None, system_prompt=None, tools=None):
+        self.calls += 1
+        self.tools_seen.append(tools)
+        assert tools is None, f"纯 JSON 模式不应下发 FC tools，实际 {tools!r}"
+        if self.calls == 1:
+            return ('好的，我来创建文件。\n'
+                    '{"action": "use_tool", "tool": "code_editor", '
+                    '"args": "create %s print(1)"}' % self._fp)
+        return "已完成。"
+
+
+class TestEndToEndDispatch:
+    def test_use_tool_dispatched_without_fc(self):
+        from xcli_core.cli_core import AICLI
+        from xcli_core.safety import get_safety, MODE_UNRESTRICTED
+
+        cli = AICLI()
+        engine = FakeEngine()
+        cli.current_engine = engine
+        cli.tui_output_callback = lambda *a, **k: None   # 同步 TUI 分支，免动画线程
+        get_safety().set_mode(MODE_UNRESTRICTED)          # 跳过交互确认，验证分发本身
+
+        d = tempfile.mkdtemp(prefix="pj_")
+        FakeEngine._fp = os.path.join(d, "demo.py")       # 原始反斜杠路径
+        fp = FakeEngine._fp
+
+        cli.process_conversation("帮我创建一个 python 文件")
+
+        assert all(t is None for t in engine.tools_seen), "引擎不应收到任何 FC tools"
+        assert engine.calls >= 2, "至少应有：首轮JSON + 工具结果续轮"
+        assert os.path.exists(fp), "code_editor 应被按名分发执行并创建文件"
+        assert "print(1)" in open(fp, encoding="utf-8").read()