Creator App Tests / PR Spam Gate (push) Skipped
Frontend Tests / PR Spam Gate (push) Skipped
NPM Format (website & console) / PR Spam Gate (push) Skipped
Pre-commit Checks / PR Spam Gate (push) Skipped
QwenPaw Creator Plugin Release / check-version (push) Canceled after 0s
QwenPaw Creator Plugin Release / upload-oss (push) Canceled after 0s
Tests / PR Spam Gate (push) Skipped
CodeQL / CodeQL (python) (push) Canceled after 0s
CodeQL / CodeQL (typescript) (push) Canceled after 0s
Creator App Tests / Creator Backend Tests (push) Canceled after 0s
Creator App Tests / Creator UI Build & Tests (push) Canceled after 0s
E2E Smoke Tests / Playwright UI Smoke (push) Canceled after 0s
Frontend Tests / Detect console changes (push) Canceled after 0s
NPM Format (website & console) / website format check (push) Canceled after 0s
NPM Format (website & console) / console format check (push) Canceled after 0s
Pre-commit Checks / Pre-commit (push) Canceled after 0s
QwenPaw Data Plugin Release / check-version (push) Canceled after 0s
Tests / Detect code changes (push) Canceled after 0s
Frontend Tests / Vitest Unit Tests (push) Canceled after 0s
Frontend Tests / Frontend Summary (push) Canceled after 0s
QwenPaw Data Plugin Release / upload-oss (push) Canceled after 0s
Tests / Maintainer Approval (push) Canceled after 0s
Tests / Unit Tests - py3.11 - macos-latest (push) Canceled after 0s
Tests / Unit Tests - py3.11 - ubuntu-latest (push) Canceled after 0s
Tests / Unit Tests - py3.13 - ubuntu-latest (push) Canceled after 0s
Tests / Unit Tests - py3.11 - windows-latest (push) Canceled after 0s
Tests / Contract Tests - py3.11 - macos-latest (push) Canceled after 0s
Tests / Contract Tests - py3.11 - ubuntu-latest (push) Canceled after 0s
Tests / Contract Tests - py3.13 - ubuntu-latest (push) Canceled after 0s
Tests / Contract Tests - py3.11 - windows-latest (push) Canceled after 0s
Tests / Integrated Tests - py3.11 - macos-latest - p0 (push) Canceled after 0s
Tests / Integrated Tests - py3.11 - macos-latest - p1 (push) Canceled after 0s
Tests / Integrated Tests - py3.11 - macos-latest - p2 (push) Canceled after 0s
Tests / Integrated Tests - py3.11 - ubuntu-latest - fallback (push) Canceled after 0s
Tests / Integrated Tests - py3.11 - ubuntu-latest - p0 (push) Canceled after 0s
Tests / Integrated Tests - py3.11 - ubuntu-latest - p1 (push) Canceled after 0s
Tests / Integrated Tests - py3.11 - ubuntu-latest - p2 (push) Canceled after 0s
Tests / Integrated Tests - py3.13 - ubuntu-latest - p0 (push) Canceled after 0s
Tests / Integrated Tests - py3.13 - ubuntu-latest - p1 (push) Canceled after 0s
Tests / Integrated Tests - py3.13 - ubuntu-latest - p2 (push) Canceled after 0s
Tests / Integrated Tests - py3.11 - windows-latest - p0 (push) Canceled after 0s
Tests / Integrated Tests - py3.11 - windows-latest - p1 (push) Canceled after 0s
Tests / Integrated Tests - py3.11 - windows-latest - p2 (push) Canceled after 0s
Tests / Coverage Report (push) Canceled after 0s
Tests / Test Summary (push) Canceled after 0s
50 lines
1.7 KiB
Python
50 lines
1.7 KiB
Python
# -*- coding: utf-8 -*-
|
|
"""The unsandboxed-recall escape hatch is gated by deployment, not agent.json.
|
|
|
|
Running recall without a sandbox executes model-authored Python as the agent
|
|
user. ``scroll_unsandboxed_allowed`` must require BOTH the per-agent config
|
|
flag AND the deployment-layer ``QWENPAW_ALLOW_UNSANDBOXED_RECALL`` env var, so
|
|
an untrusted agent.json / API payload can never bypass the sandbox on its own.
|
|
"""
|
|
|
|
from types import SimpleNamespace
|
|
|
|
import pytest
|
|
|
|
from qwenpaw.agents.context import scroll_unsandboxed_allowed
|
|
|
|
_ENV = "QWENPAW_ALLOW_UNSANDBOXED_RECALL"
|
|
|
|
|
|
def _cfg(flag: bool) -> SimpleNamespace:
|
|
return SimpleNamespace(allow_unsandboxed=flag)
|
|
|
|
|
|
def test_agent_flag_alone_is_not_enough(monkeypatch):
|
|
# The attack: agent.json sets allow_unsandboxed=true with no env opt-in.
|
|
monkeypatch.delenv(_ENV, raising=False)
|
|
assert scroll_unsandboxed_allowed(_cfg(True)) is False
|
|
|
|
|
|
def test_env_alone_is_not_enough(monkeypatch):
|
|
# Deployment allows it, but no agent requested it — stays sandboxed.
|
|
monkeypatch.setenv(_ENV, "1")
|
|
assert scroll_unsandboxed_allowed(_cfg(False)) is False
|
|
|
|
|
|
def test_both_required_grants(monkeypatch):
|
|
monkeypatch.setenv(_ENV, "1")
|
|
assert scroll_unsandboxed_allowed(_cfg(True)) is True
|
|
|
|
|
|
@pytest.mark.parametrize("val", ["true", "TRUE", "Yes", "on", " 1 "])
|
|
def test_truthy_env_values(monkeypatch, val):
|
|
monkeypatch.setenv(_ENV, val)
|
|
assert scroll_unsandboxed_allowed(_cfg(True)) is True
|
|
|
|
|
|
@pytest.mark.parametrize("val", ["", "0", "false", "no", "off", "nope"])
|
|
def test_non_truthy_env_values_deny(monkeypatch, val):
|
|
monkeypatch.setenv(_ENV, val)
|
|
assert scroll_unsandboxed_allowed(_cfg(True)) is False
|