multi-agent-mux/tests/test_issue5_phase4_runtime.py

"""Issue #5 Phase 4 — Immutable Runtime Verification & Freeze Removal.

Verifies:
  - P4-01: Eager import completeness, bytecode inhibition (PYTHONDONTWRITEBYTECODE=1),
           and MAM_VERSION protocol sync.
  - P4-02: Mid-flight on-disk code modification immunity (B-13 bug prevention).
  - P4-03: Freeze bypass under Python mode (default) and legacy fallback preservation.
"""

from __future__ import annotations

import glob
import os
import subprocess
import sys
import tempfile
from pathlib import Path
import pytest

REPO_ROOT = Path(__file__).resolve().parent.parent
SKILLS = REPO_ROOT / ".agents" / "skills"
if str(SKILLS) not in sys.path:
    sys.path.insert(0, str(SKILLS))

from lib_py.loop import eager, parser, checkpoint, runner

# Issue #17: shells started here inherit a WORKSPACE_ROOT of this test's own
# (lib.sh would otherwise rewrite the checkout's .mam/shim/herdr).
pytestmark = pytest.mark.usefixtures("own_workspace_root")


def test_p4_01_eager_import_completeness_and_bytecode_inhibition(monkeypatch):
    """P4-01: Verify all core framework modules are eagerly loaded and bytecode writes disabled."""
    loaded = eager.eager_load()
    assert eager.is_eager_loaded()
    for mod in eager.CORE_MODULES:
        assert mod in sys.modules, f"Module {mod} was not eagerly loaded into sys.modules"

    # Verify bytecode inhibition
    assert sys.dont_write_bytecode is True
    assert os.environ.get("PYTHONDONTWRITEBYTECODE") == "1"

    # Verify framework version constant sync. The literal value is not pinned
    # here: tests/test_version_consistency.py owns the lockstep between
    # eager.FRAMEWORK_VERSION, lib.sh's MAM_VERSION, VERSIONS.md and the eight
    # SKILL.md files, so a release bump must not have to edit this file too.
    eager.verify_framework_version(eager.FRAMEWORK_VERSION)

    # Verify mismatch raises RuntimeError
    monkeypatch.setenv("MAM_VERSION", "9.9.9")
    with pytest.raises(RuntimeError, match="MAM_VERSION mismatch"):
        eager.verify_framework_version()


def test_p4_01_loop_runner_instantiation_triggers_eager_load():
    """P4-01: LoopRunner.__init__ must eagerly load all modules immediately."""
    lr = runner.LoopRunner(["--creator", "test-agent", "--task", "test"])
    assert eager.is_eager_loaded()
    assert lr.creator == ""  # Not parsed until parse_args()


def test_p4_02_midflight_code_replacement_stress(tmp_path):
    """P4-02: Simulates B-13 mid-flight disk corruption.
    
    A running Python loop process that has eagerly loaded its modules must continue
    executing correctly across iterations even if underlying source files on disk
    are overwritten with syntax-breaking garbage.
    """
    # Create a sandbox directory with a copy of a worker module
    sandbox = tmp_path / "sandbox"
    sandbox.mkdir()
    mod_file = sandbox / "dummy_worker.py"
    mod_file.write_text(
        "def compute_verdict(val: int) -> str:\n"
        "    return f'PASS (score={val})'\n",
        encoding="utf-8"
    )

    runner_script = tmp_path / "run_subproc.py"
    script_content = (
        "import sys, os\n"
        f"sys.path.insert(0, r'{SKILLS}')\n"
        f"sys.path.insert(0, r'{sandbox}')\n"
        "from lib_py.loop import eager, parser, checkpoint\n"
        "import dummy_worker\n"
        "eager.eager_load()\n"
        "print('READY', flush=True)\n"
        "line = sys.stdin.readline()\n"
        "# Invoke functions after disk file has been corrupted\n"
        "res1 = dummy_worker.compute_verdict(42)\n"
        "verdict = parser.parse_verdict('Review output\\nVERDICT: PASS')\n"
        "chk_path = sys.stdin.readline().strip()\n"
        "checkpoint.save_checkpoint(chk_path, {'status': 'OK', 'res': res1, 'verdict': verdict})\n"
        "loaded = checkpoint.load_checkpoint(chk_path)\n"
        "assert loaded['status'] == 'OK'\n"
        "print('FINISHED_SUCCESSFULLY', flush=True)\n"
    )
    runner_script.write_text(script_content, encoding="utf-8")

    py_bin = sys.executable
    proc = subprocess.Popen(
        [py_bin, str(runner_script)],
        stdin=subprocess.PIPE,
        stdout=subprocess.PIPE,
        stderr=subprocess.PIPE,
        text=True,
    )

    # Wait for process to eagerly load and signal READY
    ready_line = proc.stdout.readline()
    assert ready_line.strip() == "READY", f"Child failed to start: {proc.stderr.read()}"

    # --- B-13 Corruption Phase ---
    # While process is suspended waiting on stdin, corrupt dummy_worker.py on disk
    mod_file.write_text("INVALID SYNTAX !!!\n unexpected EOF def broken():\n", encoding="utf-8")

    # Tell child process to proceed
    chk_target = tmp_path / "checkpoint.json"
    proc.stdin.write(f"PROCEED\n{chk_target}\n")
    proc.stdin.flush()

    out_line = proc.stdout.readline()
    assert out_line.strip() == "FINISHED_SUCCESSFULLY", f"Child error: {proc.stderr.read()}"

    proc.communicate(timeout=5)
    assert proc.returncode == 0
    assert chk_target.exists()


def test_p4_03_freeze_bypass_by_default(tmp_path):
    """P4-03: Under MAM_LOOP_PY=1 (default), run_loop.sh must bypass /tmp/mam-loop-freeze.*."""
    run_loop_sh = SKILLS / "multi-agent-mux-loop" / "scripts" / "run_loop.sh"
    env = os.environ.copy()
    env.pop("MAM_LOOP_FREEZE_DIR", None)
    env.pop("MAM_LOOP_FREEZE", None)
    env.pop("MAM_LOOP_PY", None)  # defaults to 1

    # Snapshot current freeze dirs in /tmp
    before_dirs = set(glob.glob(f"{tempfile.gettempdir()}/mam-loop-freeze.*"))

    res = subprocess.run(
        ["bash", str(run_loop_sh), "--help"],
        capture_output=True,
        text=True,
        env=env,
        cwd=str(REPO_ROOT),
    )
    assert res.returncode == 1
    assert "Usage: " in res.stdout

    after_dirs = set(glob.glob(f"{tempfile.gettempdir()}/mam-loop-freeze.*"))
    new_dirs = after_dirs - before_dirs
    assert len(new_dirs) == 0, f"Freeze dir was unexpectedly created: {new_dirs}"


def test_p4_03_explicit_freeze_fallback(tmp_path):
    """P4-03: Setting MAM_LOOP_FREEZE=1 or MAM_LOOP_PY=0 restores freeze snapshot logic."""
    run_loop_sh = SKILLS / "multi-agent-mux-loop" / "scripts" / "run_loop.sh"
    
    # 1. Test MAM_LOOP_FREEZE=1
    env_freeze = os.environ.copy()
    env_freeze["MAM_LOOP_FREEZE"] = "1"
    env_freeze.pop("MAM_LOOP_FREEZE_DIR", None)
    
    res = subprocess.run(
        ["bash", str(run_loop_sh), "--help"],
        capture_output=True,
        text=True,
        env=env_freeze,
        cwd=str(REPO_ROOT),
    )
    assert res.returncode == 1
    assert "Usage: " in res.stdout

    # 2. Test MAM_LOOP_NO_FREEZE=1 overrides MAM_LOOP_FREEZE=1
    env_no_freeze = os.environ.copy()
    env_no_freeze["MAM_LOOP_FREEZE"] = "1"
    env_no_freeze["MAM_LOOP_NO_FREEZE"] = "1"
    env_no_freeze.pop("MAM_LOOP_FREEZE_DIR", None)

    before_dirs = set(glob.glob(f"{tempfile.gettempdir()}/mam-loop-freeze.*"))
    res = subprocess.run(
        ["bash", str(run_loop_sh), "--help"],
        capture_output=True,
        text=True,
        env=env_no_freeze,
        cwd=str(REPO_ROOT),
    )
    assert res.returncode == 1
    after_dirs = set(glob.glob(f"{tempfile.gettempdir()}/mam-loop-freeze.*"))
    assert len(after_dirs - before_dirs) == 0