mirror of
https://github.com/shy3130/tick-stock-panel.git
synced 2026-11-02 10:26:13 +00:00
shy3130
3f8c0b169e
- 使用结构化输出适配现有助手工具循环，保留完整工具调用历史 - 严格验证工具名和参数，沿用现有动作确认和错误终态 - 旧 CLI 不支持结构化输出时给出升级提示 - 单独展示 Codex 生成超时，记录耗时与重试信号并避免输出提示词内容 补充结构化响应、工具循环、确认流程、非法输出及超时分类回归测试。 前端全量 389 项及构建通过，后端执行完整回归。 Signed-off-by: shy3130 <shy3130@qq.com>
237 lines
10 KiB
Python
237 lines
10 KiB
Python
"""Codex CLI 结构化工具轮次: 协议往返、非法输出拒绝、沿用动作确认。"""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
from pathlib import Path
|
|
|
|
import pytest
|
|
|
|
from app.custom.assistant import actions, chat_service, codex
|
|
from app.custom.assistant import tools as assistant_tools
|
|
from app.services import ai_provider
|
|
|
|
|
|
def _schema(name):
|
|
return {
|
|
"type": "function",
|
|
"function": {
|
|
"name": name,
|
|
"description": "本地工具",
|
|
"parameters": {"type": "object", "properties": {}},
|
|
},
|
|
}
|
|
|
|
|
|
@pytest.fixture
|
|
def cli_process(monkeypatch):
|
|
captured = {}
|
|
|
|
def run(args, prompt, env, timeout):
|
|
captured.update(args=list(args), prompt=prompt, timeout=timeout)
|
|
output_schema = Path(args[args.index("--output-schema") + 1])
|
|
captured["schema_path"] = output_schema
|
|
captured["schema"] = json.loads(output_schema.read_text(encoding="utf-8"))
|
|
output = Path(args[args.index("--output-last-message") + 1])
|
|
output.write_text(json.dumps(captured["response"], ensure_ascii=False), encoding="utf-8")
|
|
return 0, b"execution logs must not become the answer", b""
|
|
|
|
monkeypatch.setattr(ai_provider, "_codex_base_command", lambda: ["codex"])
|
|
monkeypatch.setattr(ai_provider, "_prepare_codex_home", lambda path: None)
|
|
monkeypatch.setattr(ai_provider, "_codex_process_env", lambda path: {})
|
|
monkeypatch.setattr(ai_provider, "_sweep_stale_codex_run_dirs", lambda: None)
|
|
monkeypatch.setattr(ai_provider, "current_ai_model", lambda: "test-model")
|
|
monkeypatch.setattr(ai_provider, "_run_codex_process", run)
|
|
return captured
|
|
|
|
|
|
async def test_codex_round_uses_schema_and_preserves_full_history(cli_process):
|
|
cli_process["response"] = {
|
|
"content": "先查询本地数据。",
|
|
"tool_calls": [
|
|
{"name": "list_strategies", "arguments": "{}"},
|
|
{"name": "get_capabilities", "arguments": '{"scope":"daily"}'},
|
|
],
|
|
}
|
|
messages = [
|
|
{"role": "system", "content": "只能使用本地数据"},
|
|
{"role": "user", "content": "我有哪些策略？"},
|
|
{"role": "assistant", "tool_calls": [{
|
|
"id": "previous", "type": "function",
|
|
"function": {"name": "list_strategies", "arguments": "{}"},
|
|
}]},
|
|
{"role": "tool", "tool_call_id": "previous", "content": '{"ok":true,"result":[]}'},
|
|
]
|
|
schemas = [_schema("list_strategies"), _schema("get_capabilities")]
|
|
events = [event async for event in codex.stream_codex_round(messages, schemas, timeout=12)]
|
|
|
|
assert events[0] == {"type": "text", "delta": "先查询本地数据。"}
|
|
calls = events[1]["tool_calls"]
|
|
assert [call["name"] for call in calls] == ["list_strategies", "get_capabilities"]
|
|
assert calls[0]["id"] and calls[0]["id"] != calls[1]["id"]
|
|
assert json.loads(calls[1]["arguments"]) == {"scope": "daily"}
|
|
assert events[1]["finish_reason"] == "tool_calls"
|
|
# 不能丢掉 assistant.tool_calls / tool_call_id, 否则 CLI 看不到已执行工具与结果的对应关系。
|
|
payload = json.loads(cli_process["prompt"].split("<user>\n", 1)[1].rsplit("\n</user>", 1)[0])
|
|
assert payload == {"messages": messages, "tools": schemas}
|
|
names = cli_process["schema"]["properties"]["tool_calls"]["items"]["properties"]["name"]["enum"]
|
|
assert set(names) == {"list_strategies", "get_capabilities"}
|
|
args = cli_process["args"]
|
|
assert args[args.index("--sandbox") + 1] == "read-only"
|
|
assert args[args.index("--model") + 1] == "test-model"
|
|
# 助手原有 240s 时限不能缩短 Codex 文本入口既有的 600s 预算。
|
|
assert cli_process["timeout"] == 600
|
|
assert not cli_process["schema_path"].exists() # 临时 schema 随隔离目录清理
|
|
|
|
|
|
async def test_codex_round_final_text(cli_process):
|
|
cli_process["response"] = {"content": "当前没有策略。", "tool_calls": []}
|
|
events = [event async for event in codex.stream_codex_round(
|
|
[{"role": "user", "content": "你好"}], [_schema("list_strategies")],
|
|
)]
|
|
assert events == [
|
|
{"type": "text", "delta": "当前没有策略。"},
|
|
{"type": "round_end", "tool_calls": [], "finish_reason": "stop"},
|
|
]
|
|
|
|
|
|
async def test_codex_round_preserves_longer_explicit_timeout(cli_process):
|
|
cli_process["response"] = {"content": "完成", "tool_calls": []}
|
|
_ = [event async for event in codex.stream_codex_round(
|
|
[{"role": "user", "content": "你好"}], [_schema("list_strategies")], timeout=900,
|
|
)]
|
|
assert cli_process["timeout"] == 900
|
|
|
|
|
|
async def test_codex_process_timeout_keeps_duration_and_distinct_error_kind(monkeypatch, caplog):
|
|
import subprocess
|
|
|
|
def fail(args, **kwargs):
|
|
raise subprocess.TimeoutExpired(args, kwargs["timeout"], stderr=b"still generating")
|
|
|
|
monkeypatch.setattr(ai_provider.subprocess, "run", fail)
|
|
with pytest.raises(RuntimeError) as error:
|
|
ai_provider._run_codex_process(["codex", "exec"], "private prompt", {}, 600)
|
|
event = chat_service._classify_exception(error.value)
|
|
assert event["kind"] == "timeout"
|
|
assert "Codex CLI" in event["message"] and "600" in event["message"]
|
|
assert "private prompt" not in caplog.text
|
|
assert "still generating" not in caplog.text
|
|
assert "600" in caplog.text and "timed out" in caplog.text
|
|
|
|
|
|
@pytest.mark.parametrize("message", [
|
|
"Codex CLI 调用超时(600 秒), 本轮生成已终止。",
|
|
"Codex CLI 调用超时, 请稍后重试或检查本机 Codex 登录状态",
|
|
])
|
|
async def test_codex_chat_timeout_is_not_collapsed_into_network_error(monkeypatch, message):
|
|
async def fake_cli(*args, **kwargs):
|
|
raise RuntimeError(message)
|
|
|
|
monkeypatch.setattr(ai_provider, "_run_codex_cli", fake_cli)
|
|
monkeypatch.setattr(chat_service, "ai_configured", lambda: True)
|
|
monkeypatch.setattr(chat_service, "is_codex_cli_provider", lambda: True)
|
|
events = [json.loads(line) async for line in chat_service.chat_stream(
|
|
history=[{"role": "user", "content": "分析工具结果"}],
|
|
)]
|
|
assert events[0]["kind"] == "timeout"
|
|
assert events[0]["message"] == message
|
|
assert events[-1]["type"] == "done"
|
|
|
|
|
|
def test_connection_failure_still_uses_network_error():
|
|
event = chat_service._classify_exception(ConnectionError("connection failed"))
|
|
assert event["kind"] == "network"
|
|
|
|
|
|
@pytest.mark.parametrize("response", [
|
|
"not JSON",
|
|
"[]",
|
|
'{"content":"答复"}',
|
|
'{"content":null,"tool_calls":[]}',
|
|
'{"content":"答复","tool_calls":{}}',
|
|
'{"content":"答复","tool_calls":[null]}',
|
|
'{"content":"答复","tool_calls":[{"name":"shell","arguments":"{}"}]}',
|
|
'{"content":"答复","tool_calls":[{"name":"list_strategies","arguments":"broken"}]}',
|
|
'{"content":"答复","tool_calls":[{"name":"list_strategies","arguments":"[]"}]}',
|
|
'{"content":"答复","tool_calls":[{"name":"list_strategies","arguments":{}}]}',
|
|
'{"content":"","tool_calls":[]}',
|
|
json.dumps({"content": "不能显示", "tool_calls": [
|
|
{"name": "list_strategies", "arguments": "{}"},
|
|
{"name": "shell", "arguments": "{}"},
|
|
]}),
|
|
])
|
|
async def test_codex_round_rejects_invalid_output_before_any_event(monkeypatch, response):
|
|
async def fake_cli(*args, **kwargs):
|
|
return response
|
|
|
|
monkeypatch.setattr(ai_provider, "_run_codex_cli", fake_cli)
|
|
events = []
|
|
with pytest.raises(RuntimeError, match="Codex CLI"):
|
|
async for event in codex.stream_codex_round([], [_schema("list_strategies")]):
|
|
events.append(event)
|
|
assert events == []
|
|
|
|
|
|
async def test_codex_round_old_cli_reports_upgrade_instead_of_text_fallback(monkeypatch):
|
|
async def fake_cli(*args, **kwargs):
|
|
raise RuntimeError("Codex CLI 调用失败: unexpected argument '--output-schema' found")
|
|
|
|
monkeypatch.setattr(ai_provider, "_run_codex_cli", fake_cli)
|
|
with pytest.raises(RuntimeError, match=r"升级.*Codex CLI"):
|
|
_ = [event async for event in codex.stream_codex_round([], [_schema("list_strategies")])]
|
|
|
|
|
|
@pytest.mark.parametrize("approve", [True, False])
|
|
async def test_codex_tools_still_use_action_confirmation(monkeypatch, approve):
|
|
outputs = [
|
|
{"content": "", "tool_calls": [{"name": "add_to_watchlist", "arguments": '{"symbol":"600519.SH"}'}]},
|
|
{"content": "完成" if approve else "已取消", "tool_calls": []},
|
|
]
|
|
requests = []
|
|
executed = []
|
|
|
|
async def fake_cli(messages, **kwargs):
|
|
requests.append(json.loads(messages[-1]["content"]))
|
|
return json.dumps(outputs.pop(0))
|
|
|
|
async def execute(name, args, ctx):
|
|
executed.append((name, args))
|
|
return {"ok": True, "result": {"added": True}}
|
|
|
|
monkeypatch.setattr(ai_provider, "_run_codex_cli", fake_cli)
|
|
monkeypatch.setattr(chat_service, "ai_configured", lambda: True)
|
|
monkeypatch.setattr(chat_service, "is_codex_cli_provider", lambda: True)
|
|
monkeypatch.setattr(actions, "auto_approve_enabled", lambda: False)
|
|
monkeypatch.setattr(assistant_tools, "execute_assistant_tool", execute)
|
|
events = []
|
|
async for line in chat_service.chat_stream(history=[{"role": "user", "content": "加入自选"}]):
|
|
event = json.loads(line)
|
|
events.append(event)
|
|
if event["type"] == "action_confirm":
|
|
assert executed == []
|
|
await actions.registry.resolve(event["call_id"], approve)
|
|
|
|
assert [event["type"] for event in events] == [
|
|
"tool_call", "action_confirm", "tool_result", "delta", "done",
|
|
]
|
|
assert executed == ([("add_to_watchlist", {"symbol": "600519.SH"})] if approve else [])
|
|
result = requests[1]["messages"][-1]
|
|
assert result["role"] == "tool"
|
|
assert json.loads(result["content"])["ok"] is approve
|
|
if not approve:
|
|
assert "拒绝" in json.loads(result["content"])["error"]
|
|
|
|
|
|
async def test_codex_failure_becomes_error_and_done(monkeypatch):
|
|
async def fake_cli(*args, **kwargs):
|
|
raise RuntimeError("Codex CLI 调用失败: 401 Unauthorized")
|
|
|
|
monkeypatch.setattr(ai_provider, "_run_codex_cli", fake_cli)
|
|
monkeypatch.setattr(chat_service, "ai_configured", lambda: True)
|
|
monkeypatch.setattr(chat_service, "is_codex_cli_provider", lambda: True)
|
|
events = [json.loads(line) async for line in chat_service.chat_stream(
|
|
history=[{"role": "user", "content": "hi"}],
|
|
)]
|
|
assert [event["type"] for event in events] == ["error", "done"]
|
|
assert "401" in events[0]["message"]
|