#!/bin/sh # Telem bootstrap — fetches the installer and hands off to it. # # --------------------------------------------------------------------------- # WHAT THIS IS # # curl -fsSL https://docs.telem.ai/install.sh | sh # (https://docs.telem.ai/alpha_install.sh serves the same bytes — the alpha-era # name, kept so links already in the wild keep working) # -> exec npm create @telemai # the installer, which does the real work # # It carries no credentials. The @telemai packages are public, so npm needs no # auth to fetch them and this script writes nothing to ~/.npmrc. (It used to: while # the packages were private, an alpha read token rode in a hosted copy of this file # and was written into the user's npm config. Both are gone — the token bought # nothing once the packages went public, and a credential left in someone's global # npm config is a cost they keep paying.) # # Everything interactive lives in the installer, not here — the rustup-init.sh # shape: a bootstrap that acquires a runtime and gets out of the way. # # --------------------------------------------------------------------------- # SUPPORTED ENVIRONMENTS # # This is a POSIX sh script and it must be RUN by a POSIX shell. Verified matrix: # # OS shell(s) sed # macOS /bin/sh (bash-posix), dash, zsh BSD sed # Linux dash (Debian/Ubuntu `sh`), bash GNU sed # Alpine / CI busybox ash busybox sed # Windows (WSL, the distro's / MSYS2's sh GNU / busybox sed # Git Bash, MSYS2) # # NODE — offered, never forced, never sudo. When node is missing or too old this # script ASKS (on /dev/tty, the `curl | sh` idiom) whether to install it, and on yes # installs a USER-LEVEL version manager into $HOME (fnm, nvm as a fallback) and the # pinned node major, then continues into the installer in the same run. It never # touches the system node and never runs a package manager with sudo. Declining (or # having no terminal to ask on, e.g. CI) prints how to install node and re-run — it # is a guided exit, not a silent one. This is the ONLY interactive step the # bootstrap carries; uv/python are handled the same way by the installer once node # exists. # # NOT SUPPORTED — native Windows `cmd.exe` and PowerShell. They cannot execute POSIX # shell syntax at all, so this script does not "half-run" there; it simply is not # theirs to run. A Windows user has two clean paths: # * run this script inside WSL or Git Bash (a real POSIX shell), OR # * skip the bootstrap and run `npm create @telemai` directly — the same # installer this script hands to. # --------------------------------------------------------------------------- set -eu TELEM_MIN_NODE=20 telem_die() { printf '%s\n' "$*" >&2 exit 1 } telem_note() { printf '%s\n' "$*" >&2 } # Translate a raw failure into an actionable message — the POSIX twin of the wizard's # translateInstallError (absorbed from the wizard's translate_error). The two failures the # The one npm failure worth translating: a bare E404 says nothing a user can act # on, so name the two causes that actually produce it. telem_translate_error() { telem_te_context="$1" telem_te_raw="${2:-}" case "$telem_te_context" in npm) case "$telem_te_raw" in *E404* | *"404 Not Found"*) printf '%s\n' "npm could not find the package. The @telemai packages are public on registry.npmjs.org, so this is normally either a typo in the package name or version, or an npm registry= setting pointing at a mirror that does not proxy npmjs." ;; *) printf 'npm failed: %s\n' "$telem_te_raw" ;; esac ;; wsl-npm) printf '%s\n' "npm resolves to a Windows install (${telem_te_raw}). Under WSL, PATH interop puts Windows npm first, and plugins would install into a Windows profile the Linux-side agent cannot load. Install Node inside your WSL distro, or remove the Windows entries from PATH." ;; *) printf '%s\n' "$telem_te_raw" ;; esac } # --- node: detect, and (new) offer a user-level install when missing/old -------- # # Node is handled the SAME way the wizard handles uv/python: detect → # explain → ASK "install it for you?" → yes: a USER-LEVEL manager in $HOME (fnm, nvm # as a fallback), never sudo, never the system node, then continue into the wizard in # the SAME run; no (or no terminal to ask on): print how to install it and re-run. # Ask a yes/no question. yes -> 0, no -> 1. The answer is read from /dev/tty because # under `curl | sh` stdin is the pipe, not the keyboard. It NEVER hangs: with no # controlling terminal (CI, a pipe with no tty) it answers "no". Tests bypass the # terminal entirely with TELEM_ASSUME_YES / TELEM_ASSUME_NO so the ask arm runs with # no tty and no network. # # The /dev/tty open is probed inside a SUBSHELL first — `( exec /dev/null || return 1 printf '%s [Y/n] ' "$1" >&2 telem_reply='' read telem_reply = TELEM_MIN_NODE is on PATH; otherwise set telem_node_problem # to a one-line explanation and return non-zero. It has no side effects, so it is safe # to call again after an install attempt to re-check. telem_node_ok() { telem_node_problem='' if ! command -v node >/dev/null 2>&1; then telem_node_problem="node is required (>= ${TELEM_MIN_NODE}) and was not found on PATH." return 1 fi telem_node_version=$(node -v 2>/dev/null || echo "") telem_node_major=$(printf '%s' "$telem_node_version" | sed -e 's/^v//' -e 's/[^0-9].*$//') case "$telem_node_major" in '' | *[!0-9]*) telem_node_problem="could not read a version from \`node -v\` (got '${telem_node_version}')." return 1 ;; esac if [ "$telem_node_major" -lt "$TELEM_MIN_NODE" ]; then telem_node_problem="node ${TELEM_MIN_NODE} or newer is required; found ${telem_node_version}." return 1 fi return 0 } # Bootstrap a user-level node manager and make `node` resolvable in THIS process. # Every step is best-effort: the post-install telem_node_ok is the real gate, so a # missing fnm or a failed `fnm install` degrades to the guidance path rather than # aborting mid-script under `set -e`. TELEM_NODE_INSTALL_CMD overrides the network # bootstrap so the multi-sh + docker matrix can exercise this arm with a stub that # drops a fake node on PATH — no network. Unset (the real default), it runs fnm. telem_install_node() { telem_note "installing node ${TELEM_MIN_NODE} at user level (no sudo)…" if [ -n "${TELEM_NODE_INSTALL_CMD:-}" ]; then eval "$TELEM_NODE_INSTALL_CMD" || telem_note "the node install command exited non-zero." else if ! (curl -fsSL https://fnm.vercel.app/install | bash) 1>&2; then telem_note "fnm bootstrap failed; trying nvm…" (curl -fsSL https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.1/install.sh | bash) 1>&2 || : fi fi telem_activate_node } # Put a freshly-installed manager's node on PATH for the rest of THIS run, so the # wizard starts without the user reopening their shell. Best-effort and silent when # no manager is present — e.g. the test stub, which already dropped node on PATH, so # neither branch fires and telem_node_ok simply finds it. telem_activate_node() { telem_fnm_dir="${XDG_DATA_HOME:-${HOME:-}/.local/share}/fnm" if [ -x "${telem_fnm_dir}/fnm" ]; then PATH="${telem_fnm_dir}:${PATH}" export PATH FNM_DIR="${telem_fnm_dir}" export FNM_DIR fnm install "${TELEM_MIN_NODE}" 1>&2 || : eval "$(fnm env 2>/dev/null)" || : fnm use "${TELEM_MIN_NODE}" 1>&2 || : fi telem_nvm_dir="${NVM_DIR:-${HOME:-}/.nvm}" if [ -s "${telem_nvm_dir}/nvm.sh" ]; then NVM_DIR="${telem_nvm_dir}" export NVM_DIR . "${telem_nvm_dir}/nvm.sh" >/dev/null 2>&1 || : nvm install "${TELEM_MIN_NODE}" >/dev/null 2>&1 || : nvm use "${TELEM_MIN_NODE}" >/dev/null 2>&1 || : fi return 0 } telem_require_node() { if ! telem_node_ok; then telem_note "$telem_node_problem" if telem_ask "Install node ${TELEM_MIN_NODE} for you? (user-level, no sudo)"; then telem_install_node if ! telem_node_ok; then telem_die "node ${TELEM_MIN_NODE} or newer is still not available after the install attempt. Install it yourself from https://nodejs.org/ or with a user-level manager: curl -fsSL https://fnm.vercel.app/install | bash && fnm install ${TELEM_MIN_NODE} then re-run this script." fi telem_note "node ${telem_node_version} is ready; continuing." else telem_die "node ${TELEM_MIN_NODE} or newer is required. Install it from https://nodejs.org/ or with a user-level manager like fnm: curl -fsSL https://fnm.vercel.app/install | bash && fnm install ${TELEM_MIN_NODE} then re-run this script." fi fi if ! command -v npm >/dev/null 2>&1; then telem_die "npm is required and was not found on PATH (it normally ships with node)." fi } # WSL PATH-interop guard (absorbed from the wizard's wsl-npm case). Under WSL, PATH interop # can put the Windows npm.exe first, so `npm` resolves to a /mnt// mount (or a # native Windows path). Plugins would then install into a Windows profile the # Linux-side agent cannot load — a dead end that reads like a plain install. Name it # and stop before writing anything or handing off. On a normal Linux/macOS npm the # path matches none of these and this is a silent no-op. telem_check_wsl_npm() { telem_npm_path=$(command -v npm 2>/dev/null || echo "") case "$telem_npm_path" in */mnt/[A-Za-z]/* | *\\*) telem_die "$(telem_translate_error wsl-npm "$telem_npm_path")" ;; esac } # Non-interactive handoff. Reached only when there is NO controlling terminal at all # (CI, a bare pipe): the interactive paths `exec` so clack inherits the real terminal, # and an exec'd process's output cannot be inspected. Here there is no clack UI to # preserve, so we CAPTURE npm's output — a 404 fetching the wizard package is a name # or version that does not exist, or a registry that does not proxy npmjs, and the raw # resolver dump says neither. We print npm's # own output, append the translation when it 404s, and exit with npm's status. The # `&& ok || status=$?` idiom keeps a non-zero npm from tripping `set -e` on the capture. telem_handoff() { telem_ho_out=$(npm create @telemai@latest -- "$@" 2>&1) && telem_ho_status=0 || telem_ho_status=$? printf '%s\n' "$telem_ho_out" if [ "$telem_ho_status" -ne 0 ]; then case "$telem_ho_out" in *E404* | *"404 Not Found"*) telem_note "" telem_note "$(telem_translate_error npm "$telem_ho_out")" ;; esac fi return "$telem_ho_status" } telem_require_node telem_check_wsl_npm # Hand off to the wizard and do no prompting of our own — the rustup-init.sh # pattern: the shell script is a bootstrap, all interactivity lives in the richer # runtime. `npm create @telemai@latest` resolves to `@telemai/create@latest` # (verified against the registry), and `--` passes our arguments through to it. # # curl | sh gives this script a PIPE as stdin, which the wizard would correctly # read as "not a terminal" and drop into flags-only mode. When a controlling # terminal exists we hand the wizard THAT instead, so the interactive path works # through the one-liner; when it does not (CI), we fall through and the wizard's # own no-TTY guard takes over. # # The test is an actual OPEN of /dev/tty, not `[ -r /dev/tty ]`: the device node # exists and its permission bits look readable even in a process with no # controlling terminal, where opening it fails with ENXIO. Testing the bits would # make every CI run die on a failed redirect instead of falling through. # # The open is done in a SUBSHELL — `( exec /dev/null` nor being inside an `if` condition contains it. A subshell scopes # that fatal exit to the subshell, so a missing /dev/tty reads here as a plain # non-zero (false) and we fall through instead of the whole bootstrap dying. bash # is lenient about this, which is exactly why the brace-group form passed under # `sh`=bash and died under `sh`=dash. if [ ! -t 0 ] && ( exec /dev/null; then exec npm create @telemai@latest -- "$@"