Teknium
383 lines
18 KiB
Python
383 lines
18 KiB
Python
"""Runtime inventory + update plan for the fleet-update pipeline.
|
|
|
|
One read-only pass answering, BEFORE any mutation: which Hermes runtimes run on this machine, how
|
|
each is deployed, which ones this update touches, and how each restarts. Every collector is a
|
|
side-effect-free probe, so ``hermes update --plan`` is safe on a live fleet.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import logging
|
|
from contextlib import contextmanager, suppress
|
|
from dataclasses import dataclass, field, asdict
|
|
from typing import Any, Optional
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
@dataclass
|
|
class RuntimeRecord:
|
|
"""One running (or expected) Hermes runtime on this machine."""
|
|
|
|
kind: str # gateway | dashboard | serve
|
|
profile: str # profile name ("default", ...)
|
|
pid: Optional[int] = None # live PID when known
|
|
supervisor: str = "manual" # systemd | launchd | desktop | manual
|
|
code_sha: Optional[str] = None # stamped running-code sha
|
|
code_version: Optional[str] = None
|
|
restart_via: str = "" # human-readable restart mechanism
|
|
detail: dict = field(default_factory=dict)
|
|
|
|
def to_dict(self) -> dict[str, Any]:
|
|
return asdict(self)
|
|
|
|
|
|
@dataclass
|
|
class UpdatePlan:
|
|
"""The full pre-update picture: install shape + runtimes + actions."""
|
|
|
|
install_method: str = "unknown" # git | docker | nix | apt | ...
|
|
updatable_in_place: bool = True
|
|
update_mechanism: str = "hermes update"
|
|
expected_sha: Optional[str] = None # current checkout HEAD (pre-pull)
|
|
expected_version: Optional[str] = None
|
|
profiles: list = field(default_factory=list)
|
|
runtimes: list = field(default_factory=list) # list[RuntimeRecord]
|
|
|
|
def to_dict(self) -> dict[str, Any]:
|
|
return asdict(self) # recursive: RuntimeRecord entries become dicts
|
|
|
|
|
|
def _detect_supervisor_for_pid(pid: int, service_pids: set, windows_service_pids: set | None = None) -> str:
|
|
"""Classify how a live gateway PID is supervised."""
|
|
if windows_service_pids and pid in windows_service_pids:
|
|
# SCM-supervised Windows gateway (WinSW/NSSM/sc.exe create): the update pause machinery
|
|
# stops the SERVICE via sc.exe instead of killing the child, so reconciliation must plan
|
|
# it under its own mechanism id, not "manual".
|
|
return "windows-service"
|
|
if pid not in service_pids:
|
|
return "manual"
|
|
with suppress(Exception):
|
|
from hermes_cli.gateway import is_macos, supports_systemd_services
|
|
|
|
if supports_systemd_services():
|
|
return "systemd"
|
|
if is_macos():
|
|
return "launchd"
|
|
return "service"
|
|
|
|
|
|
# THE restart policy table: restart execution consumes these ids via match_runtime_outcomes / the
|
|
# update's restart phase, and the receipt records per-runtime outcomes against them. Display
|
|
# strings are derived by describe_restart_mechanism — never the other way around.
|
|
_RESTART_MECHANISMS = {
|
|
"systemd": "systemd", "launchd": "launchd", "desktop": "desktop",
|
|
"windows-service": "windows-service", "manual-serve": "respawn-argv",
|
|
}
|
|
|
|
_MECHANISM_DESCRIPTIONS = {
|
|
"systemd": "systemctl restart (drain-first SIGUSR1 when supported)",
|
|
"launchd": "launchctl kickstart -k (drain-first, per-label domain)",
|
|
"desktop": "Desktop app respawns its serve backend",
|
|
"windows-service": "sc.exe stop before venv mutation, sc.exe start after update",
|
|
"respawn-argv": "stop before code swap, relaunch with recorded launch args",
|
|
}
|
|
|
|
_SERVE_KINDS = ("serve", "dashboard")
|
|
|
|
|
|
def _restart_mechanism(supervisor: str, profile: str) -> str:
|
|
"""Machine-readable restart mechanism id for a runtime."""
|
|
return _RESTART_MECHANISMS.get(supervisor, "manual")
|
|
|
|
|
|
def describe_restart_mechanism(mechanism: str, profile: str) -> str:
|
|
"""Human-readable description of a restart mechanism id."""
|
|
return _MECHANISM_DESCRIPTIONS.get(mechanism) or (
|
|
f"hermes -p {profile} gateway restart" if profile != "default" else "hermes gateway restart"
|
|
)
|
|
|
|
|
|
def _runtime(
|
|
kind: str, profile: str, pid: Optional[int], supervisor: str,
|
|
code_sha: Any = None, code_version: Any = None, **extra: Any,
|
|
) -> RuntimeRecord:
|
|
"""A :class:`RuntimeRecord` with ``restart_via`` derived from its supervisor."""
|
|
return RuntimeRecord(
|
|
kind=kind, profile=profile, pid=pid, supervisor=supervisor,
|
|
code_sha=str(code_sha) if code_sha else None, code_version=code_version,
|
|
restart_via=_restart_mechanism(supervisor, profile), **extra,
|
|
)
|
|
|
|
|
|
@contextmanager
|
|
def _probe(label: str):
|
|
"""Run one inventory collector; a failure is logged at debug and yields fewer rows, never an exception."""
|
|
try:
|
|
yield
|
|
except Exception as exc:
|
|
logger.debug("%s failed: %s", label, exc)
|
|
|
|
|
|
def collect_runtime_inventory() -> UpdatePlan:
|
|
"""Build the pre-update plan. Read-only; never raises.
|
|
|
|
Every collector degrades independently — a probe failure yields fewer rows, not an exception.
|
|
The result is embeddable in the update receipt and printable via :func:`print_update_plan`.
|
|
"""
|
|
plan = UpdatePlan()
|
|
with _probe("Install-method probe"):
|
|
from hermes_cli.config import detect_install_method, get_managed_system, recommended_update_command_for_method
|
|
|
|
method = detect_install_method()
|
|
plan.install_method = method
|
|
managed = get_managed_system()
|
|
if managed:
|
|
plan.install_method = managed
|
|
plan.updatable_in_place = method in ("git", "unknown") and not managed
|
|
# Baked image provenance: when the image marker is present it is authoritative — a
|
|
# bind-mounted checkout inside a container can look like `git` to the heuristics while
|
|
# the running filesystem is actually an immutable image. Fail-closed: an invalid marker
|
|
# still flips the plan to not-updatable.
|
|
with _probe("Image provenance probe"):
|
|
from hermes_cli.image_provenance import read_image_provenance
|
|
|
|
provenance = read_image_provenance()
|
|
if provenance is not None:
|
|
plan.updatable_in_place = False
|
|
if provenance.valid and provenance.manager:
|
|
plan.install_method = provenance.manager
|
|
plan.update_mechanism = recommended_update_command_for_method(method)
|
|
with _probe("Code-identity probe"):
|
|
from hermes_cli.build_info import get_code_identity
|
|
|
|
identity = get_code_identity(refresh=True)
|
|
plan.expected_sha = identity.get("sha")
|
|
plan.expected_version = identity.get("version")
|
|
profile_homes = []
|
|
with _probe("Profile enumeration"):
|
|
from hermes_cli.update_receipt import _profile_homes
|
|
|
|
profile_homes = _profile_homes()
|
|
plan.profiles = [name for name, _ in profile_homes]
|
|
service_pids: set = set()
|
|
with _probe("Service-PID probe"):
|
|
from hermes_cli.gateway import _get_service_pids
|
|
|
|
service_pids = _get_service_pids(all_profiles=True) or set()
|
|
|
|
# Windows SCM services (no-op off Windows): the update's pause phase stops these via `sc.exe
|
|
# stop` / restarts via `sc.exe start`, so the plan must carry the matching mechanism id.
|
|
windows_service_pids: set = set()
|
|
with _probe("Windows SCM service-ownership probe"):
|
|
from hermes_cli.gateway import find_windows_gateway_services
|
|
|
|
windows_service_pids = {int(service.gateway_pid) for service in find_windows_gateway_services()}
|
|
|
|
def _supervisor(pid: int) -> str:
|
|
return _detect_supervisor_for_pid(pid, service_pids, windows_service_pids)
|
|
# Per-profile gateways: control-socket identity first (declared by the process itself,
|
|
# including supervisor provenance — no argv/PID inference), gateway_state.json fallback.
|
|
seen_pids: set[int] = set()
|
|
with _probe("Gateway-state inventory"):
|
|
from gateway.status import _pid_exists, read_runtime_status
|
|
from hermes_cli.update_receipt import _socket_identity
|
|
|
|
for profile, home in profile_homes:
|
|
sock = _socket_identity(home)
|
|
if sock is not None:
|
|
sock_pid, identity = sock
|
|
if sock_pid in seen_pids:
|
|
# One multiplex gateway answers identify for several profile homes — one record per process.
|
|
continue
|
|
seen_pids.add(sock_pid)
|
|
declared = identity.get("supervisor")
|
|
supervisor = str(declared) if declared else _supervisor(sock_pid)
|
|
record: dict = identity
|
|
else:
|
|
record = read_runtime_status(home / "gateway_state.json") or {}
|
|
try:
|
|
sock_pid = int(record.get("pid"))
|
|
except (TypeError, ValueError):
|
|
continue
|
|
if not _pid_exists(sock_pid):
|
|
continue
|
|
seen_pids.add(sock_pid)
|
|
supervisor = _supervisor(sock_pid)
|
|
plan.runtimes.append(
|
|
_runtime("gateway", profile, sock_pid, supervisor, record.get("code_sha"), record.get("code_version"))
|
|
)
|
|
|
|
# PID-file mapped gateways not covered by a runtime-status record
|
|
with _probe("PID-file gateway inventory"):
|
|
from hermes_cli.gateway import find_profile_gateway_processes
|
|
|
|
for proc in find_profile_gateway_processes():
|
|
if proc.pid in seen_pids:
|
|
continue
|
|
seen_pids.add(proc.pid)
|
|
plan.runtimes.append(_runtime("gateway", proc.profile, proc.pid, _supervisor(proc.pid)))
|
|
|
|
# Serve/dashboard backends from the spawn ledger — runtimes the gateway collectors can never
|
|
# see (a manual `hermes serve --host <ip>` for a remote Desktop, a long-lived `hermes dashboard`).
|
|
# ledger_entries() live-verifies (pid, create_time) so PID reuse never fabricates a row. Desktop-
|
|
# supervised backends (spawner still alive) restart via the Desktop's own respawn, not ours.
|
|
with _probe("Serve/dashboard ledger inventory"):
|
|
from hermes_cli.process_identity import ledger_entries, spawner_is_dead
|
|
|
|
for entry in ledger_entries():
|
|
purpose = entry.get("purpose")
|
|
if purpose not in _SERVE_KINDS:
|
|
continue
|
|
pid = entry.get("pid")
|
|
if not isinstance(pid, int) or pid in seen_pids:
|
|
continue
|
|
seen_pids.add(pid)
|
|
# detail.create_time: process incarnation, not just the numeric PID — a post-update
|
|
# survivor probe comparing PIDs alone calls a NEW serve that reused the number a survivor.
|
|
plan.runtimes.append(_runtime(
|
|
str(purpose), str(entry.get("profile") or "default"), pid,
|
|
"desktop" if spawner_is_dead(entry) is False else "manual-serve",
|
|
detail={
|
|
"argv": entry.get("argv") or "", "host": entry.get("host") or "",
|
|
"port": entry.get("port"), "create_time": entry.get("create_time"),
|
|
},
|
|
))
|
|
return plan
|
|
|
|
|
|
def print_update_plan(plan: UpdatePlan) -> None:
|
|
"""Human-readable plan — what the update will touch and how."""
|
|
print("Update plan:")
|
|
install = f" Install: {plan.install_method}"
|
|
if plan.expected_version:
|
|
install += f" (v{plan.expected_version}" + (f" @ {plan.expected_sha[:8]}" if plan.expected_sha else "") + ")"
|
|
print(install)
|
|
if not plan.updatable_in_place:
|
|
print(" ⚠ This install is NOT updatable in place.")
|
|
print(f" Update via: {plan.update_mechanism}")
|
|
print(f" Profiles: {', '.join(plan.profiles) if plan.profiles else '(none found)'}")
|
|
if not plan.runtimes:
|
|
print(" Running Hermes services: none detected — code swap only.")
|
|
return
|
|
print(f" Running services to restart ({len(plan.runtimes)}):")
|
|
for runtime in plan.runtimes:
|
|
sha = f" @ {runtime.code_sha[:8]}" if runtime.code_sha else ""
|
|
print(f" • {runtime.kind} [{runtime.profile}] pid {runtime.pid} — {runtime.supervisor}{sha}")
|
|
print(f" restart: {describe_restart_mechanism(runtime.restart_via, runtime.profile)}")
|
|
|
|
|
|
def _serve_unit_matches_profile(profile: str, unit: object) -> bool:
|
|
"""Does *unit* name a ``hermes-serve*``/``hermes-dashboard*`` unit for *profile*? (OWN vocabulary;
|
|
the gateway's ``hermes-gateway*`` names never cover serve/dashboard runtimes.)"""
|
|
name = str(unit).removesuffix(".service").rsplit("/", 1)[-1]
|
|
suffix = "" if profile == "default" else f"-{profile}"
|
|
return name in {f"hermes-serve{suffix}", f"hermes-dashboard{suffix}"}
|
|
|
|
|
|
def match_runtime_outcomes(
|
|
plan: "UpdatePlan", *, restarted_services: list, relaunched_profiles: list,
|
|
externally_supervised_profiles: list, killed_pids: set, failed_units: list,
|
|
stale_serve_pids: "set | None" = None,
|
|
) -> list[dict[str, Any]]:
|
|
"""Reconcile the plan's runtimes against what the restart phase DID.
|
|
|
|
The platform restart branches each re-discover their own targets, so a runtime the plan saw can
|
|
be missed with no signal. Returns one ``{kind, profile, pid, mechanism, outcome}`` row per
|
|
planned runtime; outcome is ``restarted``, ``stopped``, ``failed`` or ``unaccounted`` (no
|
|
bookkeeping mentions it — the blind-spot tripwire). Never raises. Serve/dashboard runtimes are
|
|
reconciled in their OWN vocabulary and never borrow the gateway's outcome: with
|
|
``stale_serve_pids`` a pre-update serve whose incarnation is gone counts as ``restarted``, one
|
|
still alive is ``unaccounted``; without the probe an untouched serve stays ``unaccounted``.
|
|
"""
|
|
outcomes: list[dict[str, Any]] = []
|
|
try:
|
|
failed_set = {str(u) for u in (failed_units or [])}
|
|
restarted_set = {str(s) for s in (restarted_services or [])}
|
|
relaunched = set(relaunched_profiles or [])
|
|
external = set(externally_supervised_profiles or [])
|
|
killed = {int(p) for p in (killed_pids or set())}
|
|
stale_serves = {int(p) for p in stale_serve_pids} if stale_serve_pids is not None else None
|
|
|
|
def _gateway_names(r: RuntimeRecord, names: set) -> bool:
|
|
# The bare "hermes-gateway" unit name is gateway-specific: a serve/dashboard runtime
|
|
# that merely shares the default profile is a different process the gateway restart
|
|
# never touched, and must not borrow its outcome.
|
|
return any(
|
|
r.profile in name or (r.kind == "gateway" and r.profile == "default" and "hermes-gateway" in name)
|
|
for name in names
|
|
)
|
|
|
|
def _serve_outcome(r: RuntimeRecord) -> str:
|
|
# Serve/dashboard outcome in its OWN vocabulary — never the gateway's.
|
|
if r.pid is not None and r.pid in killed:
|
|
return "stopped"
|
|
if any(_serve_unit_matches_profile(r.profile, u) for u in failed_set):
|
|
return "failed"
|
|
if stale_serves is not None:
|
|
# Incarnation-verified: the pre-update process is gone (replaced by its unit / the
|
|
# dashboard cleanup respawn / the Desktop app) or it is still alive on pre-update code.
|
|
return "unaccounted" if r.pid in stale_serves else "restarted"
|
|
if any(_serve_unit_matches_profile(r.profile, s) for s in restarted_set):
|
|
return "restarted"
|
|
return "unaccounted"
|
|
|
|
for r in plan.runtimes:
|
|
if not isinstance(r, RuntimeRecord):
|
|
continue
|
|
if r.kind in _SERVE_KINDS:
|
|
outcome = _serve_outcome(r)
|
|
elif r.profile in relaunched or r.profile in external:
|
|
outcome = "restarted"
|
|
elif r.pid is not None and r.pid in killed:
|
|
outcome = "stopped"
|
|
elif _gateway_names(r, failed_set):
|
|
outcome = "failed"
|
|
elif _gateway_names(r, restarted_set):
|
|
outcome = "restarted"
|
|
else:
|
|
outcome = "unaccounted"
|
|
outcomes.append(
|
|
{"kind": r.kind, "profile": r.profile, "pid": r.pid, "mechanism": r.restart_via, "outcome": outcome}
|
|
)
|
|
except Exception as exc:
|
|
logger.debug("Runtime-outcome reconciliation failed: %s", exc)
|
|
return outcomes
|
|
|
|
|
|
def report_unaccounted_runtimes(outcomes: list[dict[str, Any]]) -> bool:
|
|
"""Print a loud warning for runtimes the restart phase never touched.
|
|
|
|
Returns True when at least one planned runtime is unaccounted; the caller escalates like a
|
|
STALE/DOWN fleet row (exit 1) — a promised restart silently missed is the class this phase
|
|
exists to kill.
|
|
"""
|
|
missed = [o for o in outcomes if o.get("outcome") == "unaccounted"]
|
|
if not missed:
|
|
return False
|
|
print()
|
|
print(" ⚠ Planned runtimes the restart phase never touched:")
|
|
for o in missed:
|
|
print(f" ✗ {o['kind']} [{o['profile']}] pid {o['pid']} — planned mechanism: {o['mechanism']}")
|
|
print(" Restart them manually, then verify:")
|
|
if any(o.get("kind") not in _SERVE_KINDS for o in missed):
|
|
print(" hermes gateway restart # active profile")
|
|
print(" hermes -p <profile> gateway restart # named profile")
|
|
if any(o.get("kind") in _SERVE_KINDS for o in missed):
|
|
# A serve/dashboard is not reachable by any `gateway restart` command: name the process,
|
|
# not the wrong verb.
|
|
print(" systemctl --user restart hermes-serve.service # unit-managed serve")
|
|
print(" relaunch `hermes serve` / `hermes dashboard` / the Desktop app")
|
|
return True
|
|
|
|
|
|
def record_plan_in_receipt(plan: UpdatePlan) -> None:
|
|
"""Attach the inventory to the active update receipt. Never raises."""
|
|
try:
|
|
import hermes_cli.update_receipt as ur
|
|
|
|
if ur._current is not None:
|
|
ur._current.data["plan"] = plan.to_dict()
|
|
except Exception as exc:
|
|
logger.debug("Could not record plan in receipt: %s", exc)
|