hermes-agent/hermes_cli/update_lock.py

"""Cross-process mutual exclusion for in-flight Hermes updates.

Until now only the Tauri updater published an "update in progress" marker (``UpdateMarkerGuard`` in
``apps/bootstrap-installer/src-tauri/src/update.rs``), and only the Electron desktop consumed it
(``electron/update-marker.ts``, to gate local backend startup).

This module makes that same marker the single lock for **all** update entrypoints instead of adding
a fourth mechanism. Format and location are unchanged and remain byte-compatible with the Rust and
Electron readers:
"""

from __future__ import annotations

import logging
import os
import time
from dataclasses import dataclass
from pathlib import Path

logger = logging.getLogger(__name__)

# Keep in sync with UPDATE_MARKER_MAX_AGE_MS in
# apps/desktop/electron/update-marker.ts — the same marker is read by both, and
# a shorter ceiling here would let Python steal a lock Electron still considers
# live. A full update (git pull + uv sync + desktop rebuild) is minutes.
UPDATE_MARKER_MAX_AGE_SECONDS = 20 * 60

MARKER_NAME = ".hermes-update-in-progress"

# Set by an orchestrating updater (the Tauri `hermes-setup --update` flow) to
# its own pid before spawning `hermes update` as a child stage. The parent
# holds the marker for its whole run, so without this the child refuses its
# own parent's lock and the GUI update can never complete. See update_child_env
# in apps/bootstrap-installer/src-tauri/src/update.rs — keep the name in sync.
HANDOFF_PID_ENV = "HERMES_UPDATE_HANDOFF_PID"

# Exit code meaning "another updater/instance owns this install right now".
# Already the de-facto contract: the Windows shim + venv-holder guards in
# _cmd_update_impl exit 2, and the Tauri updater matches on it
# (UPDATE_EXIT_CONCURRENT in apps/bootstrap-installer/src-tauri/src/update.rs)
# to show "Hermes is still running" instead of a generic failure. Naming it
# here keeps the concurrent-update refusal on that same understood contract.
UPDATE_EXIT_CONCURRENT = 2


def update_marker_path() -> Path:
    """Path of the shared update marker.

    Uses the *process* Hermes home (never the context-local profile override): the Rust updater
    resolves ``$HERMES_HOME`` or the platform default, and the desktop pins that same value into the
    updater's env. A profile-scoped path here would put the lock somewhere the other two owners
    never look.
    """
    from hermes_constants import get_process_hermes_home

    return get_process_hermes_home() / MARKER_NAME


def _pid_alive(pid: int) -> bool:
    """True when a process with ``pid`` currently exists.

    Delegates to :func:`gateway.status._pid_exists`, the project's existing no-kill probe. Do NOT
    hand-roll this with ``os.kill(pid, 0)``: on Windows that is not a no-op — CPython routes
    ``sig=0`` to ``GenerateConsoleCtrlEvent``, which Ctrl+C's the target's whole console process
    group (bpo-14484).

    Any pid we cannot evaluate counts as dead: a corrupt marker must not wedge the lock forever.
    """
    if pid <= 0:
        return False
    try:
        from gateway.status import _pid_exists

        return bool(_pid_exists(pid))
    except Exception as exc:
        # Import failure or an unusable pid (e.g. larger than the platform's
        # pid_t). Treat the marker as stale rather than blocking updates.
        logger.debug("Could not probe pid %s: %s", pid, exc)
        return False


def _handoff_pid() -> int | None:
    """Pid of the orchestrating updater that spawned us, if any.

    Read from :data:`HANDOFF_PID_ENV`. Malformed values count as absent — a broken handoff must fall
    back to the normal refusal, never crash.
    """
    raw = os.environ.get(HANDOFF_PID_ENV, "").strip()
    if not raw:
        return None
    try:
        pid = int(raw)
    except ValueError:
        return None
    return pid if pid > 0 else None


def _is_ancestor_pid(pid: int) -> bool:
    """True when ``pid`` is a live ancestor (parent chain) of this process.

    The orchestrating updater spawns ``hermes update`` as a (grand)child, so a live marker owned by
    one of our ancestors can only be the claim we are already running under — an unrelated
    concurrent updater is never in our parent chain.

    Never includes our own pid, and any failure counts as "not an ancestor": an unprovable ancestry
    must fall back to the normal refusal.
    """
    if pid <= 0:
        return False
    try:
        import psutil

        return any(parent.pid == pid for parent in psutil.Process().parents())
    except Exception as exc:
        logger.debug("Could not walk process ancestry for pid %s: %s", pid, exc)
        return False


@dataclass(frozen=True)
class UpdateHolder:
    """A confirmed-live update currently holding the lock."""

    pid: int
    age_seconds: float


def read_live_update(*, path: Path | None = None) -> UpdateHolder | None:
    """Return the live update holding the lock, or ``None``.

    Mirrors ``readLiveUpdateMarker`` in ``electron/update-marker.ts``: absent, unreadable,
    malformed, dead-pid, and past-the-ceiling all mean "no live update", and a stale marker file is
    deleted so it can't strand future runs. Never raises.
    """
    marker = path or update_marker_path()
    try:
        raw = marker.read_text(encoding="utf-8")
    except OSError:
        return None  # absent or unreadable => no live update

    lines = raw.splitlines()
    try:
        pid = int(lines[0].strip())
    except (IndexError, ValueError):
        pid = -1
    try:
        started_at = float(lines[1].strip())
    except (IndexError, ValueError):
        started_at = float("-inf")

    age = time.time() - started_at
    if not _pid_alive(pid) or age > UPDATE_MARKER_MAX_AGE_SECONDS:
        try:
            marker.unlink()
        except OSError:
            pass
        return None

    return UpdateHolder(pid=pid, age_seconds=age)


def describe_holder(holder: UpdateHolder) -> str:
    """One-line, user-facing explanation of who holds the update lock."""
    minutes, seconds = divmod(int(max(holder.age_seconds, 0)), 60)
    elapsed = f"{minutes}m {seconds}s" if minutes else f"{seconds}s"
    return (
        f"✗ Another Hermes update is already running (PID {holder.pid}, "
        f"started {elapsed} ago).\n"
        "\n"
        "  Two updates mutating the same checkout corrupt it: one rewrites\n"
        "  source while the other is mid-install. Wait for it to finish, or\n"
        "  close the window/dashboard tab that started it, then retry."
    )


class UpdateLock:
    """Context manager owning the shared update marker for this process.

    ``acquired`` is False when another live update holds it; callers decide between hard refusal
    (CLI/dashboard) and waiting. Release only removes the marker when *we* still own it, so a marker
    rewritten by a handoff partner (the Tauri updater writes its own pid) is never deleted from
    under its new owner.
    """

    def __init__(self, *, path: Path | None = None) -> None:
        self.path = path or update_marker_path()
        self.acquired = False
        self.holder: UpdateHolder | None = None

    def acquire(self) -> bool:
        """Claim the lock. Returns False (and sets ``holder``) if it's taken.

        A live holder whose pid matches :data:`HANDOFF_PID_ENV` — or is an ancestor of ours — is our
        own orchestrating parent (e.g. the Tauri updater staging ``hermes update``): run under ITS
        claim and leave its marker untouched on release. The ancestry path covers staged updaters
        older than the env-var export.
        """
        existing = read_live_update(path=self.path)
        if existing is not None:
            if existing.pid == _handoff_pid() or _is_ancestor_pid(existing.pid):
                return True
            self.holder = existing
            return False
        try:
            self.path.parent.mkdir(parents=True, exist_ok=True)
            self.path.write_text(
                f"{os.getpid()}\n{int(time.time())}\n", encoding="utf-8"
            )
        except OSError as exc:
            # Best-effort, exactly like the Rust guard: an unwritable marker
            # must not block the update itself (that would be a worse failure
            # than the race it prevents). Degrade to the pre-lock behavior.
            logger.debug("Could not write update marker %s: %s", self.path, exc)
            return True
        self.acquired = True
        return True

    def release(self) -> None:
        """Drop the marker if this process still owns it. Never raises."""
        if not self.acquired:
            return
        self.acquired = False
        try:
            raw = self.path.read_text(encoding="utf-8")
            owner = int(raw.splitlines()[0].strip())
        except (OSError, IndexError, ValueError):
            return
        if owner != os.getpid():
            # A handoff partner took ownership (e.g. the Tauri updater wrote
            # its own pid). Leave it alone — it's still a live update.
            return
        try:
            self.path.unlink()
        except OSError:
            pass

    def __enter__(self) -> "UpdateLock":
        self.acquire()
        return self

    def __exit__(self, *_exc) -> None:
        self.release()