Spaces:

Erinaldordso
/

Minecraft-Friends-AI-ZeroGPU

Paused

Minecraft-Friends-AI-ZeroGPU / tests /test_opencode_config.py

Use official OpenCode plan agent for Muse supervision

cd0a5b6 verified 29 days ago

3.96 kB

"""Behavioral checks for keyless supervision and evidence of completion."""
import importlib.util
import json
import os
from pathlib import Path

import pytest

spec = importlib.util.spec_from_file_location("opencode_config", Path(__file__).parents[1] / "opencode_config.py")
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)


def test_coach_uses_muse_only_without_authentication():
    # The AltoClef coach runs SOLELY on the free OpenCode Muse model: default
    # and small models are Muse, and no Mine/Qwen provider exists here, so no
    # title/small task can fall back to Mine. No key material anywhere.
    config = module.configuration()
    assert config["enabled_providers"] == ["opencode"]
    assert config["model"] == "opencode/muse-spark-1.3-contributor-free"
    assert config["small_model"] == "opencode/muse-spark-1.3-contributor-free"
    assert module.MUSE_MODEL == config["model"]
    assert "mine" not in config.get("provider", {})
    dumped = json.dumps(config)
    assert "apiKey" not in dumped
    assert "Authorization" not in dumped
    assert "qwen" not in dumped.lower() and "granite" not in dumped.lower()


def test_coach_uses_builtin_plan_agent_only():
    # Probed 2026-10-03 against Zen (keyless, isolated HOME, empty stdin):
    # the tested override shapes reproduce
    # `FreeTierError: ... can only be used from within OpenCode` (403):
    # prompt override, tools lockdown, blanket permission deny, even a
    # read-only permission shape. Bare description+mode and the built-ins
    # (build, plan) succeed. So there is no custom agent at all: the
    # built-in `plan` agent (read-only by design, official Muse system
    # prompt intact) supervises from user-level instructions.
    config = module.configuration()
    assert config["default_agent"] == "plan"
    assert "agent" not in config


def test_coach_instructions_live_at_user_level():
    message = module.coach_user_message({"connected": True, "task_active": True})
    assert "reply only with one compact json object" in message.lower()
    assert '"connected": true' in message
    assert "no tools" in message.lower()


def test_credentials_are_not_inherited(monkeypatch, tmp_path):
    for name in ["GEMINI_BOT_KEY_1", "OPENAI_API_KEY", "VOYAGER_API_KEY", "HF_TOKEN"]:
        monkeypatch.setenv(name, "private-secret")
    monkeypatch.setenv("PATH", "executable-search-path")
    env = module.child_environment(tmp_path)
    assert env["PATH"] == "executable-search-path"
    assert not any(env.get(name) for name in ["GEMINI_BOT_KEY_1", "OPENAI_API_KEY", "VOYAGER_API_KEY", "HF_TOKEN"])
    assert env["XDG_DATA_HOME"] == str(tmp_path / "data")


def event(text):
    return json.dumps({"type":"text","part":{"text":text}})


def test_accepts_fragmented_opencode_decision():
    output = "\n".join([event('{"action":"start",'), event('"reason":"Task is inactive"}')])
    assert module.parse_decision(output)["action"] == "start"


@pytest.mark.parametrize("text", ['{"action":"shell","reason":"run a command"}',
                                  'Dragon defeated!', '{}', '[]'])
def test_rejects_unsupported_or_invented_decisions(text):
    with pytest.raises(RuntimeError):
        module.parse_decision(event(text))


def test_provider_error_is_not_a_valid_decision():
    with pytest.raises(RuntimeError):
        module.parse_decision(json.dumps({"type":"error","error":{"message":"quota exceeded"}}))


@pytest.mark.parametrize("observation", [
    {"connected":False, "task_active":False},
    {"entered_end":True, "dragon_seen":False, "task_active":False},
    {"entered_end":True, "returned_from_end":True},
    {"credits_observed":True},
])
def test_disconnect_or_unloaded_dragon_never_proves_completion(observation):
    assert not module.goal_completed(observation)


def test_credits_after_entering_end_prove_completion():
    assert module.goal_completed({"entered_end":True,"credits_observed":True})

	"""Behavioral checks for keyless supervision and evidence of completion."""
	import importlib.util
	import json
	import os
	from pathlib import Path

	import pytest

	spec = importlib.util.spec_from_file_location("opencode_config", Path(__file__).parents[1] / "opencode_config.py")
	module = importlib.util.module_from_spec(spec)
	spec.loader.exec_module(module)


	def test_coach_uses_muse_only_without_authentication():
	# The AltoClef coach runs SOLELY on the free OpenCode Muse model: default
	# and small models are Muse, and no Mine/Qwen provider exists here, so no
	# title/small task can fall back to Mine. No key material anywhere.
	config = module.configuration()
	assert config["enabled_providers"] == ["opencode"]
	assert config["model"] == "opencode/muse-spark-1.3-contributor-free"
	assert config["small_model"] == "opencode/muse-spark-1.3-contributor-free"
	assert module.MUSE_MODEL == config["model"]
	assert "mine" not in config.get("provider", {})
	dumped = json.dumps(config)
	assert "apiKey" not in dumped
	assert "Authorization" not in dumped
	assert "qwen" not in dumped.lower() and "granite" not in dumped.lower()


	def test_coach_uses_builtin_plan_agent_only():
	# Probed 2026-10-03 against Zen (keyless, isolated HOME, empty stdin):
	# the tested override shapes reproduce
	# `FreeTierError: ... can only be used from within OpenCode` (403):
	# prompt override, tools lockdown, blanket permission deny, even a
	# read-only permission shape. Bare description+mode and the built-ins
	# (build, plan) succeed. So there is no custom agent at all: the
	# built-in `plan` agent (read-only by design, official Muse system
	# prompt intact) supervises from user-level instructions.
	config = module.configuration()
	assert config["default_agent"] == "plan"
	assert "agent" not in config


	def test_coach_instructions_live_at_user_level():
	message = module.coach_user_message({"connected": True, "task_active": True})
	assert "reply only with one compact json object" in message.lower()
	assert '"connected": true' in message
	assert "no tools" in message.lower()


	def test_credentials_are_not_inherited(monkeypatch, tmp_path):
	for name in ["GEMINI_BOT_KEY_1", "OPENAI_API_KEY", "VOYAGER_API_KEY", "HF_TOKEN"]:
	monkeypatch.setenv(name, "private-secret")
	monkeypatch.setenv("PATH", "executable-search-path")
	env = module.child_environment(tmp_path)
	assert env["PATH"] == "executable-search-path"
	assert not any(env.get(name) for name in ["GEMINI_BOT_KEY_1", "OPENAI_API_KEY", "VOYAGER_API_KEY", "HF_TOKEN"])
	assert env["XDG_DATA_HOME"] == str(tmp_path / "data")


	def event(text):
	return json.dumps({"type":"text","part":{"text":text}})


	def test_accepts_fragmented_opencode_decision():
	output = "\n".join([event('{"action":"start",'), event('"reason":"Task is inactive"}')])
	assert module.parse_decision(output)["action"] == "start"


	@pytest.mark.parametrize("text", ['{"action":"shell","reason":"run a command"}',
	'Dragon defeated!', '{}', '[]'])
	def test_rejects_unsupported_or_invented_decisions(text):
	with pytest.raises(RuntimeError):
	module.parse_decision(event(text))


	def test_provider_error_is_not_a_valid_decision():
	with pytest.raises(RuntimeError):
	module.parse_decision(json.dumps({"type":"error","error":{"message":"quota exceeded"}}))


	@pytest.mark.parametrize("observation", [
	{"connected":False, "task_active":False},
	{"entered_end":True, "dragon_seen":False, "task_active":False},
	{"entered_end":True, "returned_from_end":True},
	{"credits_observed":True},
	])
	def test_disconnect_or_unloaded_dragon_never_proves_completion(observation):
	assert not module.goal_completed(observation)


	def test_credits_after_entering_end_prove_completion():
	assert module.goal_completed({"entered_end":True,"credits_observed":True})