chenmingxuan
be6e8c2abd
Unify task scheduling and queue state, fix Android input and observation recovery, preserve screenshot paths, and match final review criteria accurately. Add collapsible copyable results and publish the complete 2.1.0 upgrade notes. Validation: 3189 backend tests, 326 frontend tests, 24 client tests, production build and quality gates.
317 lines
12 KiB
Python
317 lines
12 KiB
Python
# Copyright 2026 Google LLC
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
"""Tests for pixel safety-net target descriptions."""
|
|
|
|
from pathlib import Path
|
|
from io import BytesIO
|
|
from unittest.mock import MagicMock
|
|
from unittest.mock import AsyncMock
|
|
import base64
|
|
import pytest
|
|
from PIL import Image
|
|
from PIL import ImageDraw
|
|
|
|
from artemis.agents.validator import precondition_pixel as pp
|
|
from langchain_core.messages import HumanMessage, SystemMessage
|
|
|
|
_PNG = b"fake-image-bytes"
|
|
|
|
|
|
def _encode_image(image):
|
|
buffer = BytesIO()
|
|
image.save(buffer, format="PNG")
|
|
return base64.b64encode(buffer.getvalue()).decode()
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
@pytest.mark.parametrize("point", [[410, 2310], [1195, 251]])
|
|
async def test_pixel_judge_sees_device_target_in_each_resized_image(monkeypatch, point):
|
|
monkeypatch.setattr(pp, "_init_llm_and_prompt", lambda *a: (None, "prompt"))
|
|
completion = AsyncMock(
|
|
return_value={
|
|
"is_present": True,
|
|
"confidence": 0.99,
|
|
"target_kind": "control",
|
|
"target_matches": True,
|
|
}
|
|
)
|
|
monkeypatch.setattr(pp, "acomplete_structured", completion)
|
|
ctx = MagicMock()
|
|
ctx.device.device_width, ctx.device.device_height = 1280, 2856
|
|
reference = _encode_image(Image.new("RGB", (918, 2048), "white"))
|
|
session = AsyncMock()
|
|
session.screenshot_b64.return_value = _encode_image(Image.new("RGB", (640, 1428), "white"))
|
|
action = {"action": "tap", "coordinates": point.copy(), "target_text": "Input"}
|
|
|
|
verdict = await pp.validate_action_precondition_pixel(
|
|
ctx,
|
|
session,
|
|
action,
|
|
reference,
|
|
get_llm_fn=lambda: None,
|
|
)
|
|
|
|
assert verdict[0] is True
|
|
assert action["coordinates"] == point # Dispatch still uses native device pixels.
|
|
content = completion.await_args.args[1][1].content
|
|
images = [item["image_url"]["url"] for item in content if item["type"] == "image_url"]
|
|
assert len(images) == 2
|
|
for url in images:
|
|
with Image.open(BytesIO(base64.b64decode(url.split(",", 1)[1]))) as marked:
|
|
x = round(point[0] * marked.width / 1280)
|
|
y = round(point[1] * marked.height / 2856)
|
|
red, green, blue = marked.getpixel((x, y))
|
|
assert red > 180 and green < 60 and blue < 60
|
|
wrong = (min(point[0], marked.width - 1), min(point[1], marked.height - 1))
|
|
assert min(marked.getpixel(wrong)) > 200
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_negative_cache_tracks_scaled_target_neighborhood(monkeypatch):
|
|
monkeypatch.setattr(pp, "_init_llm_and_prompt", lambda *a: (None, "prompt"))
|
|
attempt = AsyncMock(
|
|
return_value=(False, pp.ValidationErrorCategory.PIXEL_TARGET_DISAPPEARED, "missing")
|
|
)
|
|
monkeypatch.setattr(pp, "_run_attempt", attempt)
|
|
ctx = MagicMock()
|
|
ctx.device.device_width, ctx.device.device_height = 1280, 2856
|
|
image = Image.new("RGB", (918, 2048), "white")
|
|
frames = [_encode_image(image)]
|
|
ImageDraw.Draw(image).rectangle((260, 1620, 330, 1700), fill="blue")
|
|
frames.append(_encode_image(image))
|
|
ImageDraw.Draw(image).rectangle((0, 0, 20, 20), fill="black")
|
|
frames.append(_encode_image(image))
|
|
for frame in frames:
|
|
await pp.validate_action_precondition_pixel(
|
|
ctx,
|
|
AsyncMock(),
|
|
{"action": "tap", "coordinates": [410, 2310], "target_description": "Input"},
|
|
frame,
|
|
get_llm_fn=lambda: None,
|
|
)
|
|
assert attempt.await_count == 2
|
|
|
|
|
|
def test_changed_orientation_does_not_silently_move_marker():
|
|
image = base64.b64decode(_encode_image(Image.new("RGB", (2048, 918))))
|
|
with pytest.raises(ValueError, match="geometry changed"):
|
|
pp._annotate_target(image, [410, 2310], (1280, 2856))
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
@pytest.mark.parametrize(
|
|
"response",
|
|
[
|
|
{"is_present": True, "confidence": 0.99, "target_kind": "surface", "target_matches": True},
|
|
{"is_present": True, "confidence": 0.99},
|
|
{"is_present": True, "confidence": 0.2, "target_kind": "control", "target_matches": True},
|
|
],
|
|
)
|
|
@pytest.mark.parametrize(
|
|
"target",
|
|
[
|
|
{"action": "tap", "target_text": "Input"},
|
|
{"action": "input_text", "target_description": "Input"},
|
|
],
|
|
)
|
|
async def test_named_input_is_never_approved_as_blank_or_unverified(response, target, monkeypatch):
|
|
monkeypatch.setattr(pp.visualization, "crop_and_annotate_target", lambda *a, **k: _PNG)
|
|
monkeypatch.setattr(pp, "acomplete_structured", AsyncMock(return_value=response))
|
|
session = AsyncMock()
|
|
session.screenshot_b64.return_value = base64.b64encode(_PNG).decode()
|
|
verdict = await pp._run_attempt(session, None, "prompt", _PNG, [1, 2], target, None, 1)
|
|
assert verdict[0] is False
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_renaming_rejected_point_cannot_dispatch_again(monkeypatch):
|
|
monkeypatch.setattr(pp, "_init_llm_and_prompt", lambda *a: (None, "prompt"))
|
|
attempt = AsyncMock(
|
|
return_value=(False, pp.ValidationErrorCategory.PIXEL_TARGET_DISAPPEARED, "No input")
|
|
)
|
|
monkeypatch.setattr(pp, "_run_attempt", attempt)
|
|
ctx = MagicMock()
|
|
image = Image.new("RGB", (400, 400), "white")
|
|
frames = []
|
|
for color in ("black", "blue"):
|
|
image.putpixel((5, 5), Image.new("RGB", (1, 1), color).getpixel((0, 0)))
|
|
encoded = BytesIO()
|
|
image.save(encoded, format="PNG")
|
|
frames.append(base64.b64encode(encoded.getvalue()).decode())
|
|
for description, frame in zip(("pairing input", "blank surface"), frames, strict=True):
|
|
verdict = await pp.validate_action_precondition_pixel(
|
|
ctx,
|
|
AsyncMock(),
|
|
{"action": "tap", "coordinates": [200, 300], "target_description": description},
|
|
frame,
|
|
get_llm_fn=lambda: None,
|
|
)
|
|
assert verdict[0] is False
|
|
assert attempt.await_count == 1
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
@pytest.mark.parametrize(
|
|
"screenshot_error", [RuntimeError("device offline"), TimeoutError("capture timeout")]
|
|
)
|
|
async def test_screenshot_failure_is_unavailable_and_preserves_cause(monkeypatch, screenshot_error):
|
|
monkeypatch.setattr(pp, "_init_llm_and_prompt", lambda *a: (None, "prompt"))
|
|
monkeypatch.setattr(pp.visualization, "crop_and_annotate_target", lambda *a, **k: _PNG)
|
|
monkeypatch.setattr(pp, "_RETRY_DELAY_SECONDS", 0)
|
|
complete = AsyncMock()
|
|
monkeypatch.setattr(pp, "acomplete_structured", complete)
|
|
session = AsyncMock()
|
|
session.screenshot_b64.side_effect = screenshot_error
|
|
ctx = MagicMock()
|
|
ctx.rejected_pixel_targets = {}
|
|
action = {"action": "tap", "coordinates": [200, 300], "target_text": "Projects"}
|
|
for _ in range(2):
|
|
passed, category, reason = await pp.validate_action_precondition_pixel(
|
|
ctx, session, action, base64.b64encode(_PNG).decode(), get_llm_fn=lambda: None
|
|
)
|
|
assert passed is False
|
|
assert category == pp.ValidationErrorCategory.PIXEL_VERIFICATION_UNAVAILABLE
|
|
assert str(screenshot_error) in reason
|
|
assert "action was not executed" in reason
|
|
assert session.screenshot_b64.await_count == 2 * pp._MAX_ATTEMPTS
|
|
assert not ctx.rejected_pixel_targets
|
|
complete.assert_not_awaited()
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_model_failure_is_unavailable_without_losing_parse_error(monkeypatch):
|
|
monkeypatch.setattr(pp, "_init_llm_and_prompt", lambda *a: (None, "prompt"))
|
|
monkeypatch.setattr(pp.visualization, "crop_and_annotate_target", lambda *a, **k: _PNG)
|
|
monkeypatch.setattr(pp, "_RETRY_DELAY_SECONDS", 0)
|
|
monkeypatch.setattr(
|
|
pp, "acomplete_structured", AsyncMock(side_effect=ValueError("invalid response"))
|
|
)
|
|
session = AsyncMock()
|
|
session.screenshot_b64.return_value = base64.b64encode(_PNG).decode()
|
|
passed, category, reason = await pp.validate_action_precondition_pixel(
|
|
MagicMock(),
|
|
session,
|
|
{"action": "tap", "coordinates": [200, 300], "target_text": "Projects"},
|
|
base64.b64encode(_PNG).decode(),
|
|
get_llm_fn=lambda: None,
|
|
)
|
|
assert passed is False
|
|
assert category == pp.ValidationErrorCategory.PIXEL_VERIFICATION_UNAVAILABLE
|
|
assert "invalid response" in reason
|
|
|
|
|
|
def _texts(messages: list) -> list[str]:
|
|
human = messages[1]
|
|
return [part["text"] for part in human.content if part["type"] == "text"]
|
|
|
|
|
|
def test_named_control_target_block_precedes_images():
|
|
item = {
|
|
"action": "tap",
|
|
"coordinates": [1001, 718],
|
|
"target_text": "进入全屏模式",
|
|
"target_resource_id": "com.google.android.youtube:id/fullscreen_button",
|
|
"target_class": "android.widget.ImageView",
|
|
"target_bounds": [922, 655, 1080, 781],
|
|
}
|
|
state = MagicMock()
|
|
state.operator_native_thinking = "identified 进入全屏模式"
|
|
state.operator_raw_thinking = None
|
|
|
|
messages = pp._build_messages("SYSTEM", _PNG, _PNG, item, state)
|
|
|
|
assert isinstance(messages[0], SystemMessage)
|
|
assert isinstance(messages[1], HumanMessage)
|
|
first_text = messages[1].content[0]["text"]
|
|
assert first_text.startswith("[Target]")
|
|
assert "Kind: specific UI control" in first_text
|
|
assert "Label: 进入全屏模式" in first_text
|
|
assert "Resource ID: com.google.android.youtube:id/fullscreen_button" in first_text
|
|
assert "Class: android.widget.ImageView" in first_text
|
|
assert "[922, 655, 1080, 781]" in first_text
|
|
assert messages[1].content[1]["text"] == "[Image 1 (Reference)]"
|
|
assert any("Original Thinking" in t for t in _texts(messages))
|
|
assert any("Action: tap" in t for t in _texts(messages))
|
|
|
|
|
|
def test_coordinates_only_target_is_labelled_as_surface():
|
|
item = {"action": "tap", "coordinates": [540, 400]}
|
|
|
|
messages = pp._build_messages("SYSTEM", _PNG, _PNG, item, state=None)
|
|
|
|
first_text = messages[1].content[0]["text"]
|
|
assert "Kind: coordinates only" in first_text
|
|
# No hit test runs any more: the block states that nothing was recorded, it
|
|
# does not claim the point was looked up in the UI hierarchy.
|
|
assert "no target metadata recorded" in first_text
|
|
assert "UI hierarchy" not in first_text
|
|
assert "Label:" not in first_text
|
|
assert "Resource ID:" not in first_text
|
|
assert not any("Original Thinking" in t for t in _texts(messages))
|
|
|
|
|
|
def test_label_alone_counts_as_named_control():
|
|
item = {"action": "tap", "coordinates": [10, 10], "target_text": "Submit"}
|
|
text = pp._describe_target(item)
|
|
assert "Kind: specific UI control" in text
|
|
assert "Label: Submit" in text
|
|
assert "Resource ID" not in text
|
|
|
|
|
|
def test_unlabelled_indexed_element_is_still_a_control():
|
|
item = {
|
|
"action": "tap",
|
|
"coordinates": [1001, 718],
|
|
"target_class": "android.widget.ImageView",
|
|
"target_bounds": [922, 655, 1080, 781],
|
|
}
|
|
text = pp._describe_target(item)
|
|
assert "Kind: specific UI control" in text
|
|
assert "unlabelled" in text
|
|
assert "Class: android.widget.ImageView" in text
|
|
assert "[922, 655, 1080, 781]" in text
|
|
|
|
|
|
def test_described_coordinate_target_is_its_own_kind():
|
|
"""The model's description is presented as its belief, not as observed metadata."""
|
|
item = {"action": "tap", "coordinates": [540, 400], "target_description": "play button"}
|
|
text = pp._describe_target(item)
|
|
assert "Kind: described target" in text
|
|
assert "Operator's description: play button" in text
|
|
assert "Kind: specific UI control" not in text
|
|
assert "Label:" not in text
|
|
|
|
|
|
def test_observed_metadata_outranks_description():
|
|
item = {
|
|
"action": "tap",
|
|
"coordinates": [10, 10],
|
|
"target_text": "Submit",
|
|
"target_description": "ignored",
|
|
}
|
|
text = pp._describe_target(item)
|
|
assert "Kind: specific UI control" in text
|
|
assert "Label: Submit" in text
|
|
assert "described target" not in text
|
|
|
|
|
|
def test_prompt_rules_are_keyed_on_target_kind():
|
|
prompt = Path(pp.__file__).with_name("pixel_safety_net.md").read_text(encoding="utf-8")
|
|
assert "`Kind: specific UI control`" in prompt
|
|
assert "`Kind: described target`" in prompt
|
|
assert "`Kind: coordinates only`" in prompt
|
|
assert "hidden" in prompt
|
|
assert "Identify what appears under the red dot" in prompt
|