feat: example init script to run humanlayer daemon

Author: Christer Barreholm

examples/init-humanlayer.sh

@@ -0,0 +1,89 @@
+#!/bin/bash
+# Init script for the --init-script flag: installs the HumanLayer CLI
+# and launches its daemon. Runs as root inside the container after
+# startup; the daemon itself is dropped to the claude user, which is
+# what the agent CLI runs as.
+#
+#   read -rs HUMANLAYER_LAUNCH_TOKEN && export HUMANLAYER_LAUNCH_TOKEN
+#   claude-podman \
+#     --init-script examples/init-humanlayer.sh \
+#     --init-env HUMANLAYER_LAUNCH_TOKEN
+#
+# The token is read from the environment, never passed on the command
+# line: an argument would be visible to every user on the host via `ps`,
+# and would land in your shell history.
+set -euo pipefail
+
+if [ -z "${HUMANLAYER_LAUNCH_TOKEN:-}" ]; then
+	echo "HUMANLAYER_LAUNCH_TOKEN is not set; pass it with --init-env" >&2
+	exit 1
+fi
+
+# Pinned for reproducible installs. Bump deliberately; "latest" would
+# silently pull whatever npm serves at container start.
+HUMANLAYER_VERSION="0.31.0"
+
+# nodejs/npm to run the CLI, gcompat/libstdc++ because HumanLayer ships
+# glibc-linked binaries and the container is musl (Alpine)
+apk add --no-cache nodejs npm gcompat libstdc++
+
+# Disable install scripts for every npm invocation in this container,
+# including any the agent runs later. Pre/post-install hooks are the
+# main way an npm package executes arbitrary code at install time.
+# Written to the global npmrc (root-owned, so the container user cannot
+# undo it), though note a user-level ~/.npmrc still overrides it.
+npm config set ignore-scripts true --location=global
+
+# Install as root into /usr/local: already on PATH, and the container
+# user cannot tamper with it. The flag is redundant given the config
+# above, but keeps this line correct on its own.
+npm install -g --ignore-scripts "@humanlayer/cli@${HUMANLAYER_VERSION}"
+
+humanlayer --version
+
+# The daemon runs as the agent does: uid 1000, the container's `claude`
+# user, which bin/claude maps to your host account via --userns=keep-id.
+# HumanLayer keeps its socket and session state under $HOME/.humanlayer,
+# so a root-owned daemon would put both where the CLI never looks — and
+# would give the approval broker root in the container for no reason.
+# Install as root, run as claude.
+DAEMON_USER=claude
+DAEMON_UID=$(id -u "$DAEMON_USER")
+if [ "$DAEMON_UID" != 1000 ]; then
+	echo "expected $DAEMON_USER to be uid 1000, got $DAEMON_UID" >&2
+	exit 1
+fi
+
+# `daemon launch` only runs in the foreground and has no detach flag, so
+# it must be backgrounded here — otherwise the init script never returns
+# and the agent CLI never starts. setsid detaches it from this exec
+# session; the process reparents to the container's init and lives as
+# long as the container does. There is no supervisor: if the daemon dies
+# mid-session it stays dead until you restart the container.
+#
+# The log may contain session details, so it belongs to the daemon user
+# and nobody else.
+LOG=/var/log/humanlayer-daemon.log
+touch "$LOG"
+chown "$DAEMON_USER" "$LOG"
+chmod 600 "$LOG"
+
+# su, not setpriv: Alpine's busybox setpriv has no --reuid. su also
+# gives the daemon claude's HOME, which is half the point of dropping to
+# it. Only HOME/SHELL/USER/LOGNAME are replaced, so the token still
+# reaches the inner shell through the environment instead of su's argv.
+setsid su "$DAEMON_USER" -c \
+	'exec humanlayer daemon launch --launch-token "$HUMANLAYER_LAUNCH_TOKEN"' \
+	>>"$LOG" 2>&1 </dev/null &
+DAEMON_PID=$!
+
+# Give it a moment, then confirm it is actually up: without this a bad
+# token just lands in the log and the session starts daemon-less.
+sleep 3
+if ! kill -0 "$DAEMON_PID" 2>/dev/null; then
+	echo "humanlayer daemon exited immediately; last log lines:" >&2
+	tail -n 20 "$LOG" >&2
+	exit 1
+fi
+
+echo "humanlayer daemon running (pid $DAEMON_PID), logging to $LOG"