multi-agent-mux/tests/test_issue5_phase2b_autocomplete.py

#!/usr/bin/env python3
"""Issue #5 Phase 2b — flag-gated auto-completion.

Characterize first, as in 2a: the interesting property here is what the system
*refuses* to do, so the refusals are pinned before anything is added that could
weaken them.

Sections
  1. X-01..X-05  the fail-closed preconditions as they stand today.
  2. finalize    the completion manifest: written by the agent, re-derived here.
  3. autocomplete the gate, and everything that must stop it.
  4. notifier    durable pending notification and restart recovery.
  5. flags       default off, evidence-backed allowlist, per-job opt-in.
"""
from __future__ import annotations

import hashlib
import json
import os
import subprocess
import sys
from pathlib import Path

import pytest

REPO_ROOT = Path(__file__).resolve().parent.parent
SKILLS = REPO_ROOT / ".agents" / "skills"
WATCHER_FIXTURES = REPO_ROOT / "tests" / "fixtures" / "issue_5" / "watcher"


def _reg(tmp_path) -> str:
    d = tmp_path / ".mam" / "jobs"
    d.mkdir(parents=True, exist_ok=True)
    return str(d)


def _make_job(registry_dir, **over):
    from lib_py.delegation import registry
    return registry.register_job(over.pop("prompt", "p"), registry_dir=registry_dir, **over)


def _load(registry_dir, job_id):
    from lib_py.delegation import mqtt_common
    return mqtt_common.load_job(job_id, registry_dir)


def _fingerprint(path: Path) -> dict:
    out = {}
    for p in sorted(Path(path).rglob("*")):
        if p.is_file():
            out[str(p.relative_to(path))] = hashlib.sha256(p.read_bytes()).hexdigest()
    return out


def _scenario(name="finalized_candidate"):
    return json.loads((WATCHER_FIXTURES / (name + ".json")).read_text(encoding="utf-8"))


@pytest.fixture(autouse=True)
def _default_flags(monkeypatch):
    """Auto-completion is off unless a test turns it on. That is the product
    default too, and every test that does not opt in is therefore also a check
    that the default really is off."""
    monkeypatch.delenv("MAM_AUTO_COMPLETE", raising=False)
    monkeypatch.delenv("MAM_AUTO_COMPLETE_BACKENDS", raising=False)
    monkeypatch.setenv("MAM_ENGINE_CAS", "1")


# ==============================================================================
# 1. X-01..X-05 — the fail-closed preconditions, before anything is added
# ==============================================================================

def test_x01_a_job_without_a_manifest_is_never_approved():
    """The property that makes unsupported agents safe without any new branch."""
    import copy
    from lib_py.loop import policy

    scenario = _scenario()
    assert policy.decide(scenario).decision == "candidate_approved"
    stripped = copy.deepcopy(scenario)
    for item in stripped["timeline"]:
        (item.get("payload") or {}).pop("report_manifest", None)
    stripped["filesystem"]["manifest_committed"] = False
    decision = policy.decide(stripped)
    assert decision.decision == "hold" and "manifest" in decision.reason


def test_x02_no_production_code_writes_a_completion_manifest_yet():
    """Pins the starting point: the protocol is entirely new in phase 2b."""
    hits = []
    for path in sorted((SKILLS / "lib_py").rglob("*.py")):
        if path.name in ("finalize.py", "autocomplete.py"):
            continue
        text = path.read_text(encoding="utf-8")
        if "report-manifest" in text:
            hits.append(path.name)
    assert not hits, "manifest writers already exist: {}".format(hits)


def test_x03_the_record_carries_iteration_not_an_attempt_id(mam_sandbox, tmp_path, monkeypatch):
    """Why auto-completion does not use the engine's expected_attempt.

    The engine compares expected_attempt against `attempt_id` falling back to
    `iteration`. A record has only the latter, and the two mean different things,
    so the attempt guard is applied where attempt identity actually lives: the
    dispatch record in the shadow store, compared against observed evidence."""
    monkeypatch.chdir(mam_sandbox)
    reg = _reg(tmp_path)
    job_id = _make_job(reg)
    record = _load(reg, job_id)
    assert "iteration" in record
    assert "attempt_id" not in record, (
        "if this ever changes, revisit the layering decision in the phase 2b report"
    )


def test_x04_policy_rejects_evidence_from_a_superseded_attempt():
    """The attempt guard that auto-completion actually relies on."""
    import copy
    from lib_py.loop import policy

    scenario = _scenario()
    superseded = copy.deepcopy(scenario)
    superseded["dispatch_context"]["attempt_id"] = 2   # evidence still says 1
    decision = policy.decide(superseded)
    assert decision.decision == "hold" and "attempt_id" in decision.reason


def test_x05_the_explicit_event_path_is_untouched_by_default(mam_sandbox, tmp_path, monkeypatch):
    """An explicitly completed job looks exactly as it always did."""
    monkeypatch.chdir(mam_sandbox)
    from lib_py.delegation import mqtt_common

    reg = _reg(tmp_path)
    job_id = _make_job(reg)
    mqtt_common.update_job_status(job_id, reg, _source="publish_event", status="running")
    mqtt_common.update_job_status(job_id, reg, _source="publish_event", status="completed")
    record = _load(reg, job_id)
    assert record["status"] == "completed"
    assert "completed_by" not in record, "an explicit completion carries no auto marker"