CIS490/orchestrator/episode.py
max 1b6c7b2f4a Collectors 2/4/5 + fleet runner + sample manifest + Tier-3 setup scripts
This is the chunk that makes "real data" actually flow on multiple
hosts in parallel. End-to-end pipe was up at 613c6fa / 2579683; now
the lab-host side has the diversity + concurrency it needs.

Collectors landed:
  collectors/qmp.py          — source 2 (oracle). Tiny synchronous QMP
                               client + row builder + run loop. Tolerates
                               older qemu without query-stats.
  collectors/guest_agent.py  — source 5 (deployable). Reads the
                               virtio-serial host-side socket, parses
                               agent JSON-lines, re-stamps to the host
                               monotonic clock, persists.
  collectors/pcap.py         — source 4 (deployable). tcpdump capture
                               + pure-Python pcap reader + 100 ms
                               netflow.jsonl bucketizer. Decodes
                               Ethernet/IPv4/TCP/UDP enough for the
                               schema in docs/data-model.md.

In-guest agent:
  vm/guest-agent/cis490_agent.py — stdlib-only Python agent. Reads
    /proc/{stat,meminfo,loadavg,net/dev,net/tcp*}, top-N RSS procs,
    thermal. Writes JSON-lines to /dev/virtio-ports/cis490.guest.agent.
  tools/build_cidata.py — embeds the agent + an OpenRC service into
    user-data so first boot of the Alpine cidata image auto-starts it.

Launchers:
  vm/launch_demo.sh / launch_target.sh — second virtio-serial port for
    the agent socket; SLOT env support so multiple VMs run without
    socket / port collisions; PORT_BASE on launch_target so multiple
    target VMs hostfwd different host ports.
  vm/setup_bridge.sh — creates host-only br-malware (10.200.0.1/24,
    no NAT). Idempotent.

Fleet:
  orchestrator/fleet.py — capacity detector (cores / RAM / load
    headroom) + concurrent-slot runner. Per-slot ENV selects the
    sample. FleetCapacity dataclass round-trips into meta.json so
    "this episode ran with 6 concurrent VMs" is auditable post-hoc.
  tools/run_fleet.py — CLI: --capacity report; --waves N runs N
    waves of (max_concurrent) episodes each, every slot with a
    different sample.
  etc/cis490-orchestrator.service — now drives the fleet runner with
    Restart=always so each invocation runs one wave and respawns,
    giving a continuous stream.

Samples:
  samples/manifest.toml — six profiles spanning the five major
    behaviour shapes. Each entry is real OR mimic (sha256 distinguishes).
  samples/manifest.py — strict TOML loader (rejects dups, unknown
    categories) + deterministic select(host_id, slot, episode_index)
    so different hosts on the network walk the catalog in different
    orders without any coordinator.

EpisodeRunner:
  orchestrator/episode.py — optional qmp_socket + guest_agent_socket
    fields on EpisodeConfig; when set, additional collector threads
    run alongside proc_qemu. EpisodeResult now carries rows_qmp +
    rows_guest counters.

Tier-3 setup automation:
  scripts/install-msfrpcd.sh — installs metasploit-framework where
    the package manager has it, generates a strong password into
    /etc/cis490/msfrpc.env, drops a hardened systemd unit bound to
    127.0.0.1:55553. After this, run_tier3_demo.py works zero-touch
    once MSFRPC_PASSWORD is sourced.
  scripts/fetch-metasploitable2.sh — accepts IMAGE_URL + IMAGE_SHA256
    from the operator (Rapid7 download is registration-walled), pulls,
    verifies, converts vmdk → qcow2, lands at vm/images/.

Tests: 82 pass (was 51). New suites:
  tests/test_qmp.py       — fake QMP server, capability handshake,
                            blockstats, async-event interleaving,
                            5-failure backoff
  tests/test_guest_agent.py — fake virtio socket, JSON-lines read +
                              re-stamp, malformed-line tolerance
  tests/test_pcap.py      — synthetic pcap with TCP/UDP/ARP frames,
                            bucketize correctness across windows
  tests/test_fleet.py     — capacity math (8-core idle / low-RAM /
                            high-load / Pi5 / 1-core box), manifest
                            selection determinism + diversity

What's queued for the next commit (already discussed in convo):
  - MSFExploitDriver v2: map sample.profile → distinct in-session
    workload so Tier-3 episodes don't all produce the same yes-loop
    envelope. Critical for ML to learn varied malware shapes.
  - Real-sample fetch from MalwareBazaar by sha256.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-30 00:02:27 -05:00

298 lines
10 KiB
Python

"""EpisodeRunner — single-episode driver.
Writes the full per-episode directory shape from ``docs/data-model.md``:
data/episodes/<ulid>/
meta.json
events.jsonl
labels.jsonl
telemetry-proc.jsonl
done.marker
Two modes:
1. **Single-phase** (no ``phase_schedule`` set) — labels the whole window
``clean``. Useful for v0 sanity tests against any pid.
2. **Scheduled** (``phase_schedule`` set on the config) — walks a list of
``(phase, duration_s)`` tuples, emitting one label per transition. An
optional ``on_phase`` callback fires at each transition, used to drive an
external load mimic (or, later, the real exploit/sample driver).
Real VM bring-up will arrive as a third mode that boots a guest, fires an
exploit, and adapts the schedule based on observed events
(``session_open``, ``sample_executed``, ...).
"""
from __future__ import annotations
import json
import logging
import os
import threading
import time
from dataclasses import dataclass, field
from datetime import datetime, timezone
from pathlib import Path
from typing import Callable
from collectors import guest_agent, proc_qemu, qmp
from .ulid import new_ulid
log = logging.getLogger("cis490.orchestrator")
SCHEMA_VERSION = 1
PhaseSchedule = list[tuple[str, float]]
OnPhase = Callable[[str], None]
@dataclass
class EpisodeConfig:
target_pid: int
duration_s: float
data_root: Path
interval_ms: int = 100
image_name: str = "(stub)"
snapshot_name: str = "(none-stub)"
episode_id: str | None = None
# When set, walk this schedule and ignore duration_s for sleep timing.
# ``duration_s`` still goes in meta.schedule for record-keeping.
phase_schedule: PhaseSchedule | None = None
# Optional: paths to QEMU sockets exposed by the launcher. When
# set, EpisodeRunner spins up additional collector threads.
qmp_socket: Path | None = None
qmp_interval_ms: int = 1000 # QMP queries are heavier than /proc reads
guest_agent_socket: Path | None = None
@dataclass
class EpisodeResult:
episode_id: str
episode_dir: Path
rows_proc: int
rows_qmp: int = 0
rows_guest: int = 0
pid_disappeared: bool = False
duration_observed_s: float = 0.0
phases_observed: list[str] = field(default_factory=list)
class EpisodeRunner:
def __init__(
self,
cfg: EpisodeConfig,
on_phase: OnPhase | None = None,
) -> None:
self.cfg = cfg
self.on_phase = on_phase
self.episode_id = cfg.episode_id or new_ulid()
self.episode_dir: Path = cfg.data_root / "episodes" / self.episode_id
# Create the dir up front so external drivers can call
# emit_event() between construction and run() — e.g. an exploit
# driver that writes a driver_setup event before the schedule
# walks. The dir is otherwise empty until run() opens files.
self.episode_dir.mkdir(parents=True, exist_ok=True)
self._t_mono_origin_ns: int = 0
self._stop = threading.Event()
# ---- public ---------------------------------------------------------
def run(self) -> EpisodeResult:
self._t_mono_origin_ns = time.monotonic_ns()
started_at_wall = datetime.now(timezone.utc).isoformat()
meta = self._initial_meta(started_at_wall)
self._write_meta(meta)
self.emit_event("snapshot_load", snapshot=self.cfg.snapshot_name)
rows_holder: dict[str, int] = {"proc": 0, "qmp": 0, "guest": 0}
def _proc_collector() -> None:
rows_holder["proc"] = proc_qemu.run_loop(
pid=self.cfg.target_pid,
output_path=self.episode_dir / "telemetry-proc.jsonl",
t_mono_origin_ns=self._t_mono_origin_ns,
interval_ms=self.cfg.interval_ms,
stop_event=self._stop,
)
def _qmp_collector() -> None:
assert self.cfg.qmp_socket is not None
rows_holder["qmp"] = qmp.run_loop(
socket_path=self.cfg.qmp_socket,
output_path=self.episode_dir / "telemetry-qmp.jsonl",
t_mono_origin_ns=self._t_mono_origin_ns,
interval_ms=self.cfg.qmp_interval_ms,
stop_event=self._stop,
)
def _guest_collector() -> None:
assert self.cfg.guest_agent_socket is not None
rows_holder["guest"] = guest_agent.run_loop(
socket_path=self.cfg.guest_agent_socket,
output_path=self.episode_dir / "telemetry-guest.jsonl",
t_mono_origin_ns=self._t_mono_origin_ns,
stop_event=self._stop,
)
threads: list[threading.Thread] = []
threads.append(threading.Thread(target=_proc_collector, daemon=True, name="proc_qemu"))
if self.cfg.qmp_socket is not None:
threads.append(threading.Thread(target=_qmp_collector, daemon=True, name="qmp"))
if self.cfg.guest_agent_socket is not None:
threads.append(threading.Thread(target=_guest_collector, daemon=True, name="guest_agent"))
for t in threads:
t.start()
phases_observed: list[str] = []
try:
if self.cfg.phase_schedule:
phases_observed = self._walk_schedule()
else:
self._emit_label(0, "clean", prev=None, reason="snapshot_loaded")
phases_observed = ["clean"]
self._stop.wait(timeout=self.cfg.duration_s)
finally:
self._stop.set()
for t in threads:
t.join(timeout=3.0)
pid_alive = _pid_alive(self.cfg.target_pid)
self.emit_event("episode_end", target_pid_alive=pid_alive)
end_mono_ns = time.monotonic_ns() - self._t_mono_origin_ns
meta["ended_at_wall"] = datetime.now(timezone.utc).isoformat()
meta["result"] = {
"phases_observed": phases_observed,
"rows_proc": rows_holder["proc"],
"rows_qmp": rows_holder["qmp"],
"rows_guest": rows_holder["guest"],
"pid_alive_at_end": pid_alive,
"duration_observed_s": end_mono_ns / 1_000_000_000,
}
self._write_meta(meta)
(self.episode_dir / "done.marker").touch()
log.info(
"episode %s complete: proc=%d qmp=%d guest=%d duration=%.2fs phases=%s",
self.episode_id,
rows_holder["proc"], rows_holder["qmp"], rows_holder["guest"],
end_mono_ns / 1e9,
phases_observed,
)
return EpisodeResult(
episode_id=self.episode_id,
episode_dir=self.episode_dir,
rows_proc=rows_holder["proc"],
rows_qmp=rows_holder["qmp"],
rows_guest=rows_holder["guest"],
pid_disappeared=not pid_alive,
duration_observed_s=end_mono_ns / 1_000_000_000,
phases_observed=phases_observed,
)
def stop(self) -> None:
self._stop.set()
# ---- internals ------------------------------------------------------
def _walk_schedule(self) -> list[str]:
observed: list[str] = []
prev: str | None = None
for phase, dur in self.cfg.phase_schedule or []:
if self._stop.is_set():
break
t_mono = time.monotonic_ns() - self._t_mono_origin_ns
self._emit_label(t_mono, phase, prev=prev, reason="scheduled")
self.emit_event("phase_transition", to=phase, prev=prev)
if self.on_phase is not None:
try:
self.on_phase(phase)
except Exception:
log.exception("on_phase callback raised; continuing")
observed.append(phase)
prev = phase
self._stop.wait(timeout=dur)
return observed
def _initial_meta(self, started_at_wall: str) -> dict:
return {
"episode_id": self.episode_id,
"schema_version": SCHEMA_VERSION,
"started_at_wall": started_at_wall,
"ended_at_wall": None,
"host_fingerprint": {
"kernel": os.uname().release,
"qemu_version": None,
},
"vm": {
"image_name": self.cfg.image_name,
"image_sha256": None,
"snapshot_name": self.cfg.snapshot_name,
"vcpus": None,
"ram_mib": None,
"target_pid": self.cfg.target_pid,
},
"exploit": None,
"sample": None,
"schedule": {
"baseline_seconds": self.cfg.duration_s,
"interval_ms": self.cfg.interval_ms,
"phase_schedule": self.cfg.phase_schedule,
},
"result": None,
}
def _write_meta(self, meta: dict) -> None:
path = self.episode_dir / "meta.json"
tmp = path.with_suffix(".json.partial")
with tmp.open("w") as f:
json.dump(meta, f, indent=2, sort_keys=True)
f.write("\n")
os.replace(tmp, path)
def emit_event(self, event: str, **extra) -> None:
"""Append a row to events.jsonl. Public so external drivers
(e.g. the MSF exploit driver) can stamp their own events with
the same monotonic clock the orchestrator is using."""
t_mono_ns = (
time.monotonic_ns() - self._t_mono_origin_ns
if self._t_mono_origin_ns
else 0
)
row = {
"t_mono_ns": t_mono_ns,
"t_wall_ns": time.time_ns(),
"event": event,
**extra,
}
with (self.episode_dir / "events.jsonl").open("a") as f:
f.write(json.dumps(row, sort_keys=True) + "\n")
def _emit_label(
self, t_mono_ns: int, phase: str, prev: str | None, reason: str
) -> None:
row = {
"t_mono_ns": t_mono_ns,
"t_wall_ns": time.time_ns(),
"phase": phase,
"prev": prev,
"reason": reason,
}
with (self.episode_dir / "labels.jsonl").open("a") as f:
f.write(json.dumps(row, sort_keys=True) + "\n")
def _pid_alive(pid: int) -> bool:
try:
os.kill(pid, 0)
return True
except ProcessLookupError:
return False
except PermissionError:
# Pid exists but we can't signal it.
return True