Explorar el Código

fix(python): challenge fresh file state in live SDK smoke

Tianyi Cui hace 1 día
padre
commit
e2b3800a7e

+ 2 - 2
python/development.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write python/development.md
-development.md: 39c47b7d2a86995eea017c77fb147c63f90412b1
-development.zh.md: fe7b4dbb2ca2e341dfb9ceb57959b7d9807f54d2
+development.md: a42a704aea524c142182eb295a53044a43fb7ea8
+development.zh.md: ccb6b262463da282f549ea12bbab3561834a9d67

+ 1 - 1
python/development.md

@@ -36,7 +36,7 @@ uv run --project python/sdk python scripts/smoke-python-runtime.py \
 
 Three scenarios compare committed expected output under `scripts/snapshots/python-sdk-single-exe/`. `minimal/model-visible.json` pins the Linux/macOS `sdk-minimal` profile's assembled system prompts, advertised tool schemas, and model-visible messages; `minimal/win-x64/model-visible.json` pins its PowerShell counterpart. A plugin that contributes an unintended system section or user message therefore fails the job, and every message the profile emits is compared. `advanced/` pins one complex process's SDK result and parent/child session logs across every target. `restart/` launches two complete SDK runtime processes against one persistence root and snapshots their isolated model histories, high-level results, and separate durable logs across every target. Rerun the owning scenario with `--update-snapshots` and review that diff before committing it.
 
-Trusted pull requests and master pushes also run `--scenario sdk-live --installed-wheel` on each selected native target. That scenario performs two tool-using turns against `https://api.deepseek.com`, verifies the created file externally, and fails when the repository secret is absent instead of self-skipping. Fork and Dependabot pull requests run the complete keyless installed-wheel path but receive no key.
+Trusted pull requests and master pushes also run `--scenario sdk-live --installed-wheel` on each selected native target. That scenario performs two tool-using turns against `https://api.deepseek.com`: it checks the created file immediately, replaces its content with a host-only random challenge, and requires the second turn to copy the changed content into a fresh receipt without modifying the source. Both turns must complete with the exact sentinel answer and a model-requested tool call; external byte comparisons check the files. Missing repository secrets fail instead of self-skipping. Fork and Dependabot pull requests run the complete keyless installed-wheel path but receive no key.
 
 An interactive smoke test needs `DEEPSEEK_API_KEY` in the environment or repository-root `.env`:
 

+ 1 - 1
python/development.zh.md

@@ -36,7 +36,7 @@ uv run --project python/sdk python scripts/smoke-python-runtime.py \
 
 其中三个场景会比对 `scripts/snapshots/python-sdk-single-exe/` 下已提交的期望输出。`minimal/model-visible.json` 固定 Linux/macOS `sdk-minimal` profile 所组装的系统提示词、对外公布的工具 schema 与模型可见消息;`minimal/win-x64/model-visible.json` 固定对应的 PowerShell 版本。因此,插件一旦贡献出计划外的系统分段或 user 消息,该任务即失败,且该 profile 发出的每条消息都会参与比对。`advanced/` 跨所有目标固定一个复杂进程的 SDK 结果及父/子会话日志。`restart/` 针对同一持久化根目录启动两个完整 SDK 运行时进程,并跨所有目标固定其彼此隔离的模型历史、高层结果与独立持久日志。重新运行对应场景时加上 `--update-snapshots`,并在提交前审阅该差异。
 
-可信拉取请求与 master 推送还会在各自选定的原生目标上运行 `--scenario sdk-live --installed-wheel`。该场景面向 `https://api.deepseek.com` 执行两个使用工具的轮次,从外部验证已创建文件,并在仓库密钥缺失时失败而不是自行 skip。Fork 与 Dependabot 拉取请求会运行完整的 keyless 安装后 wheel 路径,但不会获得密钥。
+可信拉取请求与 master 推送还会在各自选定的原生目标上运行 `--scenario sdk-live --installed-wheel`。该场景面向 `https://api.deepseek.com` 执行两个使用工具的轮次:立即检查已创建文件,将其内容替换为仅宿主知道的随机挑战值,并要求第二轮将变更后的内容复制到全新的回执文件,且不修改源文件。两个轮次都必须完成、返回精确的哨兵答案并由模型请求调用工具;文件通过外部逐字节比较验证。仓库密钥缺失时失败,而不是自行 skip。Fork 与 Dependabot 拉取请求会运行完整的 keyless 安装后 wheel 路径,但不会获得密钥。
 
 交互式冒烟测试需要环境变量或仓库根目录 `.env` 中存在 `DEEPSEEK_API_KEY`:
 

+ 143 - 0
python/sdk/tests/test_smoke_model.py

@@ -4,14 +4,157 @@ import json
 import runpy
 import subprocess
 from pathlib import Path
+from types import SimpleNamespace
 
 import pytest
 
+from deepseek_harness import RunResult
+
 
 ROOT = Path(__file__).resolve().parents[3]
 SMOKE = runpy.run_path(ROOT / "scripts" / "smoke-python-runtime.py")
 
 
+def live_result(**overrides: object) -> RunResult:
+    values = {
+        "session_id": "installed-wheel-live-api",
+        "final_response": SMOKE["LIVE_API_SENTINEL"],
+        "finish_reason": "completed",
+        "events": [{"type": "tool/call", "data": {"name": "unrelated_tool"}}],
+        "notifications": [],
+    }
+    values.update(overrides)
+    return RunResult(**values)
+
+
+@pytest.fixture
+def live_smoke(monkeypatch: pytest.MonkeyPatch) -> SimpleNamespace:
+    import deepseek_harness
+
+    state = SimpleNamespace(
+        prompts=[], session_ids=[], challenges=[], checked_logs=[], closed=False,
+        create_bytes=SMOKE["LIVE_API_SENTINEL"].encode("utf-8"),
+        create_result=live_result(), verify_result=live_result(), receipt_mode="copy",
+    )
+    globals_ = SMOKE["smoke_sdk_live"].__globals__
+    token_hex = globals_["secrets"].token_hex
+
+    def fresh_challenge(size: int) -> str:
+        assert len(state.prompts) == 1
+        value = token_hex(size)
+        state.challenges.append(value)
+        return value
+
+    class ScriptedHarness:
+        def __init__(self, **kwargs: object) -> None:
+            state.root = Path(kwargs["cwd"])
+            assert "toolChoice" not in kwargs
+
+        def __enter__(self) -> ScriptedHarness:
+            return self
+
+        def __exit__(self, *args: object) -> None:
+            state.closed = True
+
+        def run(self, prompt: str, *, session_id: str) -> RunResult:
+            state.prompts.append(prompt)
+            state.session_ids.append(session_id)
+            if len(state.prompts) == 1:
+                state.marker = Path(prompt.splitlines()[-1])
+                assert state.marker.parent == state.root
+                if state.create_bytes is not None:
+                    state.marker.write_bytes(state.create_bytes)
+                return state.create_result
+
+            assert len(state.prompts) == 2
+            assert len(state.challenges) == 1
+            challenge = state.challenges[0]
+            assert all(challenge not in sent for sent in state.prompts)
+            assert str(state.marker) not in prompt
+            assert "previous turn" in prompt and "changed externally" in prompt
+            assert state.marker.read_bytes() == challenge.encode("ascii")
+            receipt = Path(prompt.splitlines()[-1])
+            assert receipt != state.marker and not receipt.exists()
+            if state.receipt_mode == "copy":
+                receipt.write_bytes(state.marker.read_bytes())
+            elif state.receipt_mode == "stale":
+                receipt.write_bytes(state.create_bytes)
+            elif state.receipt_mode == "wrong":
+                receipt.write_bytes(b"wrong")
+            elif state.receipt_mode == "newline":
+                receipt.write_bytes(state.marker.read_bytes() + b"\n")
+            elif state.receipt_mode == "changed-source":
+                receipt.write_bytes(state.marker.read_bytes())
+                state.marker.write_bytes(b"changed")
+            elif state.receipt_mode != "missing":
+                raise AssertionError(state.receipt_mode)
+            return state.verify_result
+
+    monkeypatch.setenv("DEEPSEEK_API_KEY", "unit-test-key")
+    monkeypatch.setenv("DEEPSEEK_BASE_URL", "https://api.invalid")
+    monkeypatch.setattr(deepseek_harness, "DeepSeekHarness", ScriptedHarness)
+    monkeypatch.setattr(globals_["secrets"], "token_hex", fresh_challenge)
+    monkeypatch.setitem(globals_, "assert_zstd_session_log", state.checked_logs.append)
+    return state
+
+
+def test_live_smoke_requires_fresh_external_content(live_smoke: SimpleNamespace) -> None:
+    SMOKE["smoke_sdk_live"]()
+    assert len(live_smoke.prompts) == 2
+    assert live_smoke.session_ids == ["installed-wheel-live-api"] * 2
+    assert len(live_smoke.checked_logs) == 1
+    assert live_smoke.closed and not live_smoke.root.exists()
+
+
+@pytest.mark.parametrize("label", ["create", "verify"])
+@pytest.mark.parametrize(("overrides", "message"), [
+    ({"finish_reason": "error"}, "turn ended with 'error'"),
+    ({"events": []}, "turn made no model-requested tool call"),
+    ({"final_response": "PYTHON_SDK_LIVE_OK extra"}, "turn returned"),
+])
+def test_live_smoke_rejects_invalid_turn_before_continuing(
+    live_smoke: SimpleNamespace, label: str, overrides: dict[str, object], message: str,
+) -> None:
+    setattr(live_smoke, f"{label}_result", live_result(**overrides))
+    with pytest.raises(AssertionError, match=f"{label} {message}"):
+        SMOKE["smoke_sdk_live"]()
+    assert len(live_smoke.prompts) == (1 if label == "create" else 2)
+    assert not live_smoke.checked_logs
+    assert live_smoke.closed
+    if label == "create":
+        assert not live_smoke.challenges
+
+
+@pytest.mark.parametrize("content", [None, b"wrong", b"PYTHON_SDK_LIVE_OK\n"])
+def test_live_smoke_rejects_bad_create_before_host_overwrite(
+    live_smoke: SimpleNamespace, content: bytes | None,
+) -> None:
+    live_smoke.create_bytes = content
+    with pytest.raises(AssertionError, match="create turn (did not create|wrote unexpected bytes)"):
+        SMOKE["smoke_sdk_live"]()
+    assert len(live_smoke.prompts) == 1
+    assert not live_smoke.challenges and not live_smoke.checked_logs
+    assert live_smoke.closed
+
+
+@pytest.mark.parametrize(("mode", "message"), [
+    ("missing", "did not create receipt"),
+    ("stale", "wrote unexpected bytes to receipt"),
+    ("wrong", "wrote unexpected bytes to receipt"),
+    ("newline", "wrote unexpected bytes to receipt"),
+    ("changed-source", "changed source file"),
+])
+def test_live_smoke_rejects_unrelated_tool_without_exact_receipt(
+    live_smoke: SimpleNamespace, mode: str, message: str,
+) -> None:
+    live_smoke.receipt_mode = mode
+    with pytest.raises(AssertionError, match=f"verify turn {message}"):
+        SMOKE["smoke_sdk_live"]()
+    assert len(live_smoke.prompts) == 2
+    assert not live_smoke.checked_logs
+    assert live_smoke.closed
+
+
 @pytest.mark.parametrize(
     ("prompt_name", "expected"),
     [

+ 50 - 30
scripts/smoke-python-runtime.py

@@ -11,6 +11,7 @@ import json
 import os
 import queue
 import re
+import secrets
 import shutil
 import subprocess
 import sys
@@ -904,12 +905,9 @@ def smoke_sdk_live() -> None:
         session_id = "installed-wheel-live-api"
         shell_tool = "pwsh" if IS_WINDOWS else "bash"
         create_prompt = (
-            f"Use the {shell_tool} tool to create the file at the absolute path below with exactly one line "
-            f"containing {LIVE_API_SENTINEL}. Then reply with exactly {LIVE_API_SENTINEL}.\n{marker}"
-        )
-        verify_prompt = (
-            "Use a tool to read the file created in the previous turn. "
-            f"If its only line is {LIVE_API_SENTINEL}, reply with exactly {LIVE_API_SENTINEL}."
+            f"Use the {shell_tool} tool to create the file at the absolute path below with exact UTF-8 "
+            f"content {LIVE_API_SENTINEL}, with no newline or byte-order mark. "
+            f"Then reply with exactly {LIVE_API_SENTINEL}.\n{marker}"
         )
         with DeepSeekHarness(
             provider="deepseek-official",
@@ -925,34 +923,56 @@ def smoke_sdk_live() -> None:
             request_timeout_seconds=180,
         ) as harness:
             created = harness.run(create_prompt, session_id=session_id)
-            verified = harness.run(verify_prompt, session_id=session_id)
-
-        for label, result in (("create", created), ("verify", verified)):
-            if result.finish_reason != "completed":
-                event_types = [event.get("type") for event in result.events]
-                turn_end_data = next(
-                    (event.get("data") for event in reversed(result.events) if event.get("type") == "turn/end"),
-                    None,
-                )
-                turn_end = safe_turn_end(turn_end_data)
-                raise AssertionError(
-                    f"{label} turn ended with {result.finish_reason!r}; "
-                    f"final={result.final_response!r}; turn_end={turn_end!r}; events={event_types}"
-                )
-            if not any(event.get("type") == "tool/call" for event in result.events):
-                raise AssertionError(
-                    f"{label} turn made no model-requested tool call; "
-                    f"final={result.final_response!r}"
+            assert_live_turn("create", created)
+            if not marker.is_file():
+                raise AssertionError(f"create turn did not create {marker}")
+            if marker.read_bytes() != LIVE_API_SENTINEL.encode("utf-8"):
+                raise AssertionError(f"create turn wrote unexpected bytes to {marker}")
+
+            # The challenge is absent from the prior turn and the verification prompt.
+            challenge = secrets.token_hex(32).encode("ascii")
+            marker.write_bytes(challenge)
+            with tempfile.TemporaryDirectory(prefix="receipt-", dir=root) as receipt_directory:
+                receipt = Path(receipt_directory) / "receipt.txt"
+                verify_prompt = (
+                    "The file created in the previous turn has changed externally. "
+                    "Use a tool to read that same file and copy its exact current content to the "
+                    "new receipt path below, without changing the source file. "
+                    "Preserve every byte; do not add a newline or byte-order mark. "
+                    f"Then reply with exactly {LIVE_API_SENTINEL}.\n{receipt}"
                 )
-            if result.final_response.strip() != LIVE_API_SENTINEL:
-                raise AssertionError(f"{label} turn returned {result.final_response!r}")
-        if not marker.is_file():
-            raise AssertionError(f"real-model tool turn did not create {marker}")
-        if marker.read_text(encoding="utf-8").splitlines() != [LIVE_API_SENTINEL]:
-            raise AssertionError(f"real-model tool turn wrote unexpected text to {marker}")
+                verified = harness.run(verify_prompt, session_id=session_id)
+                assert_live_turn("verify", verified)
+                if not receipt.is_file():
+                    raise AssertionError(f"verify turn did not create receipt {receipt}")
+                if receipt.read_bytes() != challenge:
+                    raise AssertionError(f"verify turn wrote unexpected bytes to receipt {receipt}")
+                if not marker.is_file() or marker.read_bytes() != challenge:
+                    raise AssertionError(f"verify turn changed source file {marker}")
         assert_zstd_session_log(sessions)
 
 
+def assert_live_turn(label: str, result: RunResult) -> None:
+    """Require completed model tool use and the exact smoke answer for each live turn."""
+    if result.finish_reason != "completed":
+        event_types = [event.get("type") for event in result.events]
+        turn_end_data = next(
+            (event.get("data") for event in reversed(result.events) if event.get("type") == "turn/end"),
+            None,
+        )
+        turn_end = safe_turn_end(turn_end_data)
+        raise AssertionError(
+            f"{label} turn ended with {result.finish_reason!r}; "
+            f"final={result.final_response!r}; turn_end={turn_end!r}; events={event_types}"
+        )
+    if not any(event.get("type") == "tool/call" for event in result.events):
+        raise AssertionError(
+            f"{label} turn made no model-requested tool call; "
+            f"final={result.final_response!r}"
+        )
+    if result.final_response.strip() != LIVE_API_SENTINEL:
+        raise AssertionError(f"{label} turn returned {result.final_response!r}")
+
 def safe_turn_end(value: object) -> object:
     """Project a live-provider failure without retaining credential-bearing text."""
     if not isinstance(value, dict):