fix(mosaic): wire constrained recovery runtime boundaries

This commit is contained in:
wjarvis mos-comms
2026-07-19 20:03:16 -05:00
parent f4beedc3e7
commit 95681510c1
14 changed files with 687 additions and 102 deletions

View File

@@ -11,6 +11,7 @@ from concurrent.futures import ThreadPoolExecutor
import json
import os
import secrets
import select
import signal
import socket
import stat
@@ -28,7 +29,11 @@ if _MODULE_DIRECTORY not in sys.path:
sys.path.insert(0, _MODULE_DIRECTORY)
from normative_fragments import build_payload_from_wire
from receipt_challenge import is_verbatim_receipt, latest_assistant_digest, receipt_for
from receipt_observer import FileTestReceiptObserver, ReceiptObserver, UnavailableReceiptObserver
from receipt_observer import (
FileTestReceiptObserver,
ReceiptObserver,
RuntimeReceiptObserver,
)
MAX_FRAME: Final = 64 * 1024
MAX_STATE: Final = 4 * 1024 * 1024
@@ -302,7 +307,9 @@ class StateStore:
class Broker:
def __init__(self, store: StateStore, observer: ReceiptObserver | None = None) -> None:
self.store = store
self.observer: ReceiptObserver = observer if observer is not None else UnavailableReceiptObserver()
# Production construction always has a transport-capable observer. Test
# fixtures may inject their controlled observer explicitly.
self.observer: ReceiptObserver = observer if observer is not None else RuntimeReceiptObserver()
# Set only by begin_verification after its mandatory revoke-first fence.
# It is preserved if later cycle admission is refused; all other broker
# actions retain the normal snapshot rollback behavior.
@@ -390,6 +397,40 @@ class Broker:
raise BrokerFailure("STATE_INTEGRITY")
lease["state"] = LEASE_VERIFIED
def record_runtime_observation(
self, peer_pid: int, request: dict[str, object]
) -> dict[str, object]:
"""Record only an authenticated adapter's finalized assistant message.
This method is deliberately unreachable through ``Broker.handle`` and
its public broker socket. The production observer socket calls it after
SO_PEERCRED/ancestry authentication, preserving the S1 rule that a
broker request can never carry ``latest_assistant_message``.
"""
required = {
"action", "session_id", "runtime_generation", "runtime", "latest_assistant_message"
}
if set(request) != required or request.get("action") != "record_runtime_observation":
raise BrokerFailure("INVALID_OBSERVATION")
runtime = request.get("runtime")
message = request.get("latest_assistant_message")
if runtime not in READ_ONLY_TOOLS or not isinstance(message, str) or len(message.encode("utf-8")) > MAX_FRAME:
raise BrokerFailure("INVALID_OBSERVATION")
session_id, _ = self.authenticate(peer_pid, request)
generation = request["runtime_generation"]
lease = self.leases.get(session_id)
if (
not isinstance(lease, dict)
or lease.get("state") != LEASE_PENDING
or lease.get("runtime") != runtime
or lease.get("runtime_generation") != generation
or not isinstance(self.observer, RuntimeReceiptObserver)
):
raise BrokerFailure("OBSERVATION_UNAVAILABLE")
self.observer.record_latest_assistant_message(session_id, runtime, generation, message)
return {"ok": True}
def handle(self, peer: tuple[int, int, int], request: dict[str, object]) -> dict[str, object]:
if self.store.poisoned:
raise StateCommitUncertain()
@@ -783,14 +824,63 @@ def handle_connection(
return
def handle_runtime_observation_connection(
connection: socket.socket,
broker: Broker,
broker_lock: threading.Lock,
) -> None:
"""Serve the authenticated production observer transport, never the broker API."""
with connection:
read_deadline = time.monotonic() + READ_DEADLINE_SECONDS
try:
raw = connection.getsockopt(socket.SOL_SOCKET, socket.SO_PEERCRED, 12)
peer = struct.unpack("3i", raw)
request = read_frame(connection, read_deadline)
except BrokerFailure as exc:
reply = {"ok": False, "code": exc.code}
except OSError:
return
else:
acquired = broker_lock.acquire(timeout=HANDLE_QUEUE_TIMEOUT_SECONDS)
if not acquired:
reply = {"ok": False, "code": "BROKER_BUSY"}
else:
try:
try:
reply = broker.record_runtime_observation(peer[0], request)
except BrokerFailure as exc:
reply = {"ok": False, "code": exc.code}
finally:
broker_lock.release()
try:
connection.settimeout(SEND_TIMEOUT_SECONDS)
connection.sendall((json.dumps(reply, separators=(",", ":")) + "\n").encode())
except OSError:
return
def serve(
socket_path: Path,
state_path: Path,
observer: ReceiptObserver | None = None,
observer_socket_path: Path | None = None,
) -> None:
secure_parent(socket_path)
if socket_path.exists() or socket_path.is_symlink():
raise BrokerFailure("SOCKET_ALREADY_EXISTS")
runtime_observer = observer is None
if runtime_observer:
observer = RuntimeReceiptObserver()
observer_socket_path = observer_socket_path or socket_path.with_name("receipt-observer.sock")
if observer_socket_path == socket_path:
raise BrokerFailure("OBSERVER_SOCKET_CONFLICT")
secure_parent(observer_socket_path)
if observer_socket_path.exists() or observer_socket_path.is_symlink():
raise BrokerFailure("OBSERVER_SOCKET_ALREADY_EXISTS")
elif observer_socket_path is not None:
raise BrokerFailure("TEST_OBSERVER_SOCKET_CONFLICT")
store = StateStore(state_path)
broker = Broker(store, observer)
broker_lock = threading.Lock()
@@ -805,16 +895,28 @@ def serve(
server.bind(str(socket_path))
os.chmod(socket_path, 0o600)
owned = (socket_path.stat().st_dev, socket_path.stat().st_ino)
observer_server: socket.socket | None = None
observer_owned: tuple[int, int] | None = None
if runtime_observer and observer_socket_path is not None:
observer_server = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
observer_server.bind(str(observer_socket_path))
os.chmod(observer_socket_path, 0o600)
observer_owned = (observer_socket_path.stat().st_dev, observer_socket_path.stat().st_ino)
stopping = False
def stop(_signum: int, _frame: object) -> None:
nonlocal stopping
stopping = True
server.close()
if observer_server is not None:
observer_server.close()
def process_connection(connection: socket.socket) -> None:
def process_connection(connection: socket.socket, is_observer: bool) -> None:
try:
handle_connection(connection, broker, broker_lock)
if is_observer:
handle_runtime_observation_connection(connection, broker, broker_lock)
else:
handle_connection(connection, broker, broker_lock)
except Exception as exc:
with fatal_lock:
if not fatal_errors:
@@ -829,57 +931,73 @@ def serve(
signal.signal(signal.SIGTERM, stop)
signal.signal(signal.SIGINT, stop)
server.listen(MAX_IN_FLIGHT_CONNECTIONS)
server.settimeout(0.1)
server.setblocking(False)
if observer_server is not None:
observer_server.listen(MAX_IN_FLIGHT_CONNECTIONS)
observer_server.setblocking(False)
print("READY", flush=True)
try:
while not stopping:
failure = fatal_error()
if failure is not None:
raise failure
if not slots.acquire(timeout=0.1):
continue
listeners = [server, *([observer_server] if observer_server is not None else [])]
try:
connection, _ = server.accept()
except socket.timeout:
slots.release()
continue
except OSError:
slots.release()
failure = fatal_error()
if failure is not None:
raise failure
ready, _, _ = select.select(listeners, [], [], 0.1)
except (OSError, ValueError):
if stopping:
break
raise
try:
executor.submit(process_connection, connection)
except Exception:
slots.release()
connection.close()
raise
for listener in ready:
if not slots.acquire(blocking=False):
continue
try:
connection, _ = listener.accept()
except BlockingIOError:
slots.release()
continue
except OSError:
slots.release()
if stopping:
break
raise
try:
executor.submit(process_connection, connection, listener is observer_server)
except Exception:
slots.release()
connection.close()
raise
finally:
server.close()
if observer_server is not None:
observer_server.close()
executor.shutdown(wait=True)
try:
current = socket_path.stat()
if (current.st_dev, current.st_ino) == owned:
socket_path.unlink()
except FileNotFoundError:
pass
for path, inode in ((socket_path, owned), (observer_socket_path, observer_owned)):
if path is None or inode is None:
continue
try:
current = path.stat()
if (current.st_dev, current.st_ino) == inode:
path.unlink()
except FileNotFoundError:
pass
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--socket", required=True, type=Path)
parser.add_argument("--state", required=True, type=Path)
parser.add_argument("--observer-socket", type=Path)
parser.add_argument("--test-observer-file", type=Path)
arguments = parser.parse_args()
if arguments.observer_socket is not None and arguments.test_observer_file is not None:
raise BrokerFailure("TEST_OBSERVER_SOCKET_CONFLICT")
observer = (
FileTestReceiptObserver(arguments.test_observer_file)
if arguments.test_observer_file is not None
else None
)
serve(arguments.socket, arguments.state, observer)
serve(arguments.socket, arguments.state, observer, arguments.observer_socket)
if __name__ == "__main__":

View File

@@ -97,6 +97,12 @@ def main(
environment["MOSAIC_RUNTIME_GENERATION"] = str(generation)
environment["MOSAIC_LEASE_GENERATION_FILE"] = str(generation_file)
environment["MOSAIC_LEASE_RUNTIME"] = arguments.runtime
# Matches daemon.py's production default; deployments using a distinct
# observer socket may set this authenticated transport path explicitly.
environment.setdefault(
"MOSAIC_RECEIPT_OBSERVER_SOCKET",
str(socket_path.with_name("receipt-observer.sock")),
)
try:
execute(command[0], command, environment)
except OSError:

View File

@@ -8,14 +8,21 @@ import json
import os
import socket
import sys
import shlex
from collections.abc import Callable, Mapping, Sequence
from pathlib import Path
from typing import BinaryIO, Final
# Isolated (`python -I`) adapter invocations must still import co-located
# framework modules; never depend on the caller's PYTHONPATH.
_MODULE_DIRECTORY = str(Path(__file__).resolve().parent)
if _MODULE_DIRECTORY not in sys.path:
sys.path.insert(0, _MODULE_DIRECTORY)
from lease_generation import read_runtime_generation
MAX_FRAME: Final = 64 * 1024
BROKER_TIMEOUT_SECONDS: Final = 1.5
RECOVERY_TOOL: Final = "mosaic_context_recover"
def deny(code: str) -> int:
@@ -23,7 +30,7 @@ def deny(code: str) -> int:
return 2
def read_tool_name(stream: BinaryIO | None = None) -> str:
def read_tool_request(stream: BinaryIO | None = None) -> dict[str, object]:
source = sys.stdin.buffer if stream is None else stream
raw = source.read(MAX_FRAME + 1)
if len(raw) > MAX_FRAME:
@@ -34,7 +41,53 @@ def read_tool_name(stream: BinaryIO | None = None) -> str:
tool_name = value.get("tool_name")
if not isinstance(tool_name, str) or not tool_name or len(tool_name) > 256:
raise ValueError("INVALID_GATE_INPUT")
return tool_name
return value
def read_tool_name(stream: BinaryIO | None = None) -> str:
"""Backward-compatible strict extraction for callers that need only the name."""
return str(read_tool_request(stream)["tool_name"])
def recovery_invocation_name(request: dict[str, object], recovery_command: Path | None) -> str:
"""Map only a direct Claude Bash invocation of the recovery executable.
The mapping is adapter-configured and remains broker-authenticated through
the ordinary authorization request. It never blesses Bash generally, shell
composition, a different executable, or unsupported recovery arguments.
"""
tool_name = request["tool_name"]
if tool_name != "Bash" or recovery_command is None:
return str(tool_name)
tool_input = request.get("tool_input")
if not isinstance(tool_input, dict) or set(tool_input) != {"command"}:
return str(tool_name)
command = tool_input.get("command")
if not isinstance(command, str) or not command:
return str(tool_name)
try:
argv = shlex.split(command, posix=True)
except ValueError:
return str(tool_name)
if len(argv) < 3 or argv[0] != "python3":
return str(tool_name)
try:
if Path(argv[1]).resolve(strict=False) != recovery_command.resolve(strict=False):
return str(tool_name)
except OSError:
return str(tool_name)
phase = argv[2]
if phase == "complete" and len(argv) == 3:
return RECOVERY_TOOL
if phase != "begin" or len(argv) != 9:
return str(tool_name)
expected_flags = {"--construction", "--compaction-epoch", "--request-epoch"}
supplied_flags = set(argv[3::2])
if supplied_flags != expected_flags or any(not value for value in argv[4::2]):
return str(tool_name)
return RECOVERY_TOOL
def broker_request(socket_path: Path, request: dict[str, object]) -> dict[str, object]:
@@ -70,11 +123,13 @@ def main(
) -> int:
parser = argparse.ArgumentParser()
parser.add_argument("--runtime", required=True, choices=("claude", "pi"))
parser.add_argument("--recovery-command", type=Path)
arguments = parser.parse_args(argv)
source_environment = os.environ if environ is None else environ
try:
tool_name = read_tool_name(stream)
request_input = read_tool_request(stream)
tool_name = recovery_invocation_name(request_input, arguments.recovery_command)
socket_value = source_environment["MOSAIC_LEASE_BROKER_SOCKET"]
session_id = source_environment["MOSAIC_LEASE_SESSION_ID"]
generation = resolve_generation(source_environment)

View File

@@ -0,0 +1,143 @@
#!/usr/bin/env python3
"""Authenticated adapter-to-daemon transport for finalized assistant receipts.
This is not a broker request client. It writes only to the daemon-owned observer
socket, which authenticates SO_PEERCRED/ancestry before retaining a message for
the broker's ReceiptObserver seam.
"""
from __future__ import annotations
import argparse
import json
import os
import socket
import stat
import sys
from collections.abc import Mapping, Sequence
from pathlib import Path
from typing import Final
MAX_FRAME: Final = 64 * 1024
BROKER_TIMEOUT_SECONDS: Final = 1.5
MAX_TRANSCRIPT_BYTES: Final = 4 * 1024 * 1024
def read_json(stream: object) -> dict[str, object]:
raw = getattr(stream, "buffer", stream).read(MAX_FRAME + 1)
if not isinstance(raw, bytes) or len(raw) > MAX_FRAME:
raise ValueError("invalid observer input")
value = json.loads(raw)
if not isinstance(value, dict):
raise ValueError("invalid observer input")
return value
def assistant_text(entry: object) -> str | None:
if not isinstance(entry, dict):
return None
message = entry.get("message", entry)
if not isinstance(message, dict) or message.get("role") != "assistant":
return None
content = message.get("content")
if isinstance(content, str):
return content
if not isinstance(content, list):
return None
parts: list[str] = []
for item in content:
if not isinstance(item, dict) or item.get("type") != "text" or not isinstance(item.get("text"), str):
return None
parts.append(item["text"])
return "".join(parts)
def claude_latest_entry(value: dict[str, object]) -> str:
transcript_path = value.get("transcript_path")
if not isinstance(transcript_path, str) or not transcript_path:
raise ValueError("invalid Claude observer input")
path = Path(transcript_path)
flags = os.O_RDONLY | getattr(os, "O_CLOEXEC", 0) | getattr(os, "O_NOFOLLOW", 0)
descriptor = os.open(path, flags)
try:
metadata = os.fstat(descriptor)
if not stat.S_ISREG(metadata.st_mode) or metadata.st_size > MAX_TRANSCRIPT_BYTES:
raise ValueError("unsafe Claude transcript")
raw = os.read(descriptor, MAX_TRANSCRIPT_BYTES + 1)
finally:
os.close(descriptor)
if len(raw) > MAX_TRANSCRIPT_BYTES:
raise ValueError("oversized Claude transcript")
for line in reversed(raw.decode("utf-8").splitlines()):
try:
text = assistant_text(json.loads(line))
except json.JSONDecodeError as exc:
raise ValueError("invalid Claude transcript") from exc
if text is not None:
return text
raise ValueError("Claude transcript has no assistant entry")
def pi_message_end(value: dict[str, object]) -> str:
if set(value) != {"latest_assistant_message"} or not isinstance(value["latest_assistant_message"], str):
raise ValueError("invalid Pi observer input")
return value["latest_assistant_message"]
def observer_request(socket_path: Path, request: dict[str, object]) -> dict[str, object]:
payload = (json.dumps(request, separators=(",", ":")) + "\n").encode()
if len(payload) > MAX_FRAME:
raise ValueError("observer request too large")
response = bytearray()
with socket.socket(socket.AF_UNIX, socket.SOCK_STREAM) as connection:
connection.settimeout(BROKER_TIMEOUT_SECONDS)
connection.connect(str(socket_path))
connection.sendall(payload)
connection.shutdown(socket.SHUT_WR)
while len(response) <= MAX_FRAME:
chunk = connection.recv(min(4096, MAX_FRAME + 1 - len(response)))
if not chunk:
break
response.extend(chunk)
if len(response) > MAX_FRAME or not response.endswith(b"\n"):
raise ValueError("invalid observer reply")
value = json.loads(response)
if not isinstance(value, dict):
raise ValueError("invalid observer reply")
return value
def main(argv: Sequence[str] | None = None, *, environ: Mapping[str, str] | None = None) -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--runtime", required=True, choices=("claude", "pi"))
parser.add_argument("--latest-entry", action="store_true")
arguments = parser.parse_args(argv)
source_environment = os.environ if environ is None else environ
try:
source = read_json(sys.stdin)
if arguments.runtime == "claude":
if not arguments.latest_entry:
raise ValueError("Claude observer requires --latest-entry")
message = claude_latest_entry(source)
else:
if arguments.latest_entry:
raise ValueError("Pi observer is message_end only")
message = pi_message_end(source)
if len(message.encode("utf-8")) > MAX_FRAME:
raise ValueError("assistant message too large")
reply = observer_request(Path(source_environment["MOSAIC_RECEIPT_OBSERVER_SOCKET"]), {
"action": "record_runtime_observation",
"session_id": source_environment["MOSAIC_LEASE_SESSION_ID"],
"runtime_generation": int(source_environment["MOSAIC_RUNTIME_GENERATION"]),
"runtime": arguments.runtime,
"latest_assistant_message": message,
})
except (KeyError, OSError, ValueError, json.JSONDecodeError) as error:
print(f"Mosaic receipt observer refused: {error}", file=sys.stderr)
return 2
return 0 if reply == {"ok": True} else 2
if __name__ == "__main__":
raise SystemExit(main())

View File

@@ -1,9 +1,11 @@
#!/usr/bin/env python3
"""Trusted latest-assistant-message observer boundary for receipt promotion.
Runtime adapters must implement ``observe_latest_assistant_message`` directly:
Claude selects the exact latest assistant entry and Pi selects ``message_end``.
The broker accepts no observed message through its request protocol.
Production adapters deliver finalized assistant content over the daemon-owned
observer socket after the daemon authenticates their peer against the broker's
kernel-anchored session identity. The broker request protocol never accepts
assistant-message content. Claude supplies its latest assistant entry; Pi
supplies finalized assistant content at ``message_end``.
"""
from __future__ import annotations
@@ -26,7 +28,7 @@ class ReceiptObserver(Protocol):
class UnavailableReceiptObserver:
"""Production-safe default until a runtime adapter injects an observer."""
"""Fail-closed only for direct unit construction without daemon transport."""
def observe_latest_assistant_message(
self,
@@ -38,6 +40,31 @@ class UnavailableReceiptObserver:
return None
class RuntimeReceiptObserver:
"""Daemon-owned production observer populated only by authenticated adapters."""
def __init__(self) -> None:
self._messages: dict[tuple[str, str, int], str] = {}
def record_latest_assistant_message(
self,
session_id: str,
runtime: str,
runtime_generation: int,
message: str,
) -> None:
self._messages[(session_id, runtime, runtime_generation)] = message
def observe_latest_assistant_message(
self,
session_id: str,
runtime: str,
runtime_generation: int,
_binding: dict[str, object],
) -> str | None:
return self._messages.get((session_id, runtime, runtime_generation))
class TestReceiptObserver:
"""Deterministic controlled observer used only by byte-build tests."""
@@ -60,7 +87,7 @@ class TestReceiptObserver:
class FileTestReceiptObserver:
"""Private fixture-file observer for isolated out-of-process test drivers."""
"""Private fixture-file observer for isolated out-of-process test drivers only."""
def __init__(self, path: Path) -> None:
self.path = path

View File

@@ -18,6 +18,11 @@ from collections.abc import Mapping, Sequence
from pathlib import Path
from typing import Final
# The out-of-process recovery command is intentionally runnable with `python
# -I`; locate its shipped construction module without caller-controlled paths.
_MODULE_DIRECTORY = str(Path(__file__).resolve().parent)
if _MODULE_DIRECTORY not in sys.path:
sys.path.insert(0, _MODULE_DIRECTORY)
from normative_fragments import build_payload_from_wire
MAX_FRAME: Final = 64 * 1024