#!/usr/bin/env sh
# basis-worker · install.sh — one-command setup for a Basis contributor worker.
#
# WHAT THIS DOES (read-only + idempotent; it never spends, signs, or commits):
#   1. Detects OS (Linux / macOS / WSL), arch (x64 / arm64), and the GPU/runtime
#      (NVIDIA + CUDA + VRAM, Apple Metal, AMD/ROCm, or CPU fallback).
#   2. Detects Docker + the NVIDIA Container Toolkit, Ollama / vLLM, Node, Python.
#   3. Recommends a runtime + a model class for the detected hardware.
#   4. Fetches the basis-worker skill package (SKILL.md + scripts) from
#      https://basis.watch into ./basis-worker-skill so an agent can run the flow.
#   5. Prints the exact next steps: doctor → register → publish offer → start.
#
# HARD SAFETY RULES (enforced by construction):
#   - It NEVER asks for, accepts, generates, prints, or stores a PRIVATE KEY or a
#     SEED PHRASE. The reward wallet is a PUBLIC address only (0x + 40 hex), and it
#     is supplied LATER, by you, to the register step — never to this installer.
#   - It NEVER prints, writes, or transmits a worker token / auth key / secret.
#   - It writes nothing outside the current directory and makes no on-chain action.
#   - It is safe to re-run: existing files are refreshed, never destroyed wholesale.
#
# USAGE
#   curl -fsSL https://basis.watch/api/skills/basis-worker/install.sh -o install.sh
#   curl -fsSL https://basis.watch/api/skills/basis-worker/install.sh.sha256 -o install.sh.sha256
#   sha256sum -c install.sh.sha256        # verify before running (recommended)
#   sh install.sh                          # detect + fetch the skill (no install of GPU drivers)
#
# Flags:
#   --dir <path>   where to place the fetched skill (default ./basis-worker-skill)
#   --base <url>   Basis base URL (default https://basis.watch)
#   --no-fetch     detect + recommend only; do not download the skill package
#   --help         show this help
#
# This script does NOT install GPU drivers, Docker, Ollama, or vLLM for you — it
# detects them and tells you exactly what is missing and where to get it. Driver /
# toolkit installation is a deliberate, host-specific operator step.

set -eu

BASE_URL="https://basis.watch"
SKILL_DIR="./basis-worker-skill"
DO_FETCH=1

# ── tiny helpers ──────────────────────────────────────────────────────────────
say()  { printf '%s\n' "$*"; }
ok()   { printf '  \033[32m✓\033[0m %s\n' "$*"; }
no()   { printf '  \033[2m·\033[0m %s\n' "$*"; }
warn() { printf '  \033[33m!\033[0m %s\n' "$*"; }
hdr()  { printf '\n\033[1m%s\033[0m\n' "$*"; }
have() { command -v "$1" >/dev/null 2>&1; }

# ── args ──────────────────────────────────────────────────────────────────────
while [ $# -gt 0 ]; do
  case "$1" in
    --dir)   SKILL_DIR="${2:?--dir needs a path}"; shift 2 ;;
    --base)  BASE_URL="${2:?--base needs a url}"; shift 2 ;;
    --no-fetch) DO_FETCH=0; shift ;;
    --help|-h)
      sed -n '2,38p' "$0" | sed 's/^# \{0,1\}//'
      exit 0 ;;
    *) warn "ignoring unknown argument: $1"; shift ;;
  esac
done
BASE_URL="$(printf '%s' "$BASE_URL" | sed 's#/$##')"

say ""
say "  basis-worker · installer"
say "  ────────────────────────"
say "  Turn this machine into a Basis contributor inference worker."
say "  Reward wallet = a PUBLIC address only. This installer handles NO key/seed."

# ── 1. OS + arch ──────────────────────────────────────────────────────────────
hdr "1. Host"
UNAME_S="$(uname -s 2>/dev/null || echo unknown)"
UNAME_M="$(uname -m 2>/dev/null || echo unknown)"
OS="unknown"; IS_WSL=0
case "$UNAME_S" in
  Linux)  OS="linux";  grep -qi microsoft /proc/version 2>/dev/null && IS_WSL=1 ;;
  Darwin) OS="macos" ;;
  CYGWIN*|MINGW*|MSYS*) OS="windows" ;;
esac
ARCH="other"
case "$UNAME_M" in
  x86_64|amd64) ARCH="x64" ;;
  arm64|aarch64) ARCH="arm64" ;;
esac
if [ "$IS_WSL" = 1 ]; then ok "OS: Linux (WSL) / $ARCH"; else ok "OS: $OS / $ARCH"; fi

# ── 2. Node + Python ──────────────────────────────────────────────────────────
hdr "2. Runtimes"
NODE_OK=0
if have node; then
  NODE_V="$(node -p 'process.versions.node' 2>/dev/null || echo '?')"
  NODE_MAJOR="$(printf '%s' "$NODE_V" | cut -d. -f1)"
  if [ "${NODE_MAJOR:-0}" -ge 20 ] 2>/dev/null; then ok "Node $NODE_V"; NODE_OK=1
  else warn "Node $NODE_V is too old — the worker tooling needs Node >= 20 (https://nodejs.org)"; fi
else
  no "Node not found — install Node >= 20 (https://nodejs.org) to run the skill scripts"
fi
if have python3; then ok "Python $(python3 -V 2>&1 | awk '{print $2}')"; else no "python3 not found (only needed for a vLLM pip install)"; fi

# ── 3. GPU / accelerator ──────────────────────────────────────────────────────
hdr "3. Accelerator"
GPU_CLASS="cpu"        # cpu | nvidia | apple | amd
VRAM_GB=0
if have nvidia-smi; then
  GPU_NAME="$(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | head -n1 || echo '')"
  VRAM_MIB="$(nvidia-smi --query-gpu=memory.total --format=csv,noheader,nounits 2>/dev/null | head -n1 || echo 0)"
  case "$VRAM_MIB" in ''|*[!0-9]*) VRAM_MIB=0 ;; esac
  VRAM_GB=$(( VRAM_MIB / 1024 ))
  CUDA_V="$(nvidia-smi 2>/dev/null | sed -n 's/.*CUDA Version: \([0-9.]*\).*/\1/p' | head -n1)"
  GPU_CLASS="nvidia"
  ok "NVIDIA GPU: ${GPU_NAME:-detected} (${VRAM_GB} GB VRAM${CUDA_V:+, CUDA $CUDA_V})"
elif [ "$OS" = "macos" ] && [ "$ARCH" = "arm64" ]; then
  GPU_CLASS="apple"
  ok "Apple Silicon (Metal) — run Ollama natively (no --gpus all)"
elif have rocminfo || have rocm-smi; then
  GPU_CLASS="amd"
  ok "AMD GPU (ROCm) detected — run Ollama natively; vLLM-ROCm is advanced"
else
  no "No NVIDIA/Apple/AMD accelerator detected — the GPU-free 'echo' runtime still works (CPU)"
fi

# ── 4. Containers + model runtimes ────────────────────────────────────────────
hdr "4. Containers + model runtimes"
NVIDIA_RT=0
if have docker; then
  ok "Docker: $(docker --version 2>/dev/null | sed 's/,.*//')"
  if docker info --format '{{.Runtimes}}' 2>/dev/null | grep -qi nvidia; then
    NVIDIA_RT=1; ok "NVIDIA Container Toolkit present (docker run --gpus all works)"
  elif [ "$GPU_CLASS" = "nvidia" ]; then
    warn "NVIDIA Container Toolkit NOT found — needed for 'docker run --gpus all'"
    warn "  install: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html"
  fi
else
  no "Docker not found (optional — only needed for the containerized Ollama/vLLM path)"
fi
if have ollama; then ok "Ollama: $(ollama --version 2>/dev/null | head -n1)"; else no "Ollama not installed (recommended real-inference runtime — https://ollama.com)"; fi
if have vllm || python3 -c 'import vllm' >/dev/null 2>&1; then ok "vLLM available"; else no "vLLM not detected (optional, high-throughput NVIDIA serving)"; fi

# ── 5. Recommendation ─────────────────────────────────────────────────────────
hdr "5. Recommendation"
RUNTIME="echo"; MODEL_CLASS="(none — echo simulator)"; MODEL_HINT="prove the pipeline with no model first"
if [ "$GPU_CLASS" = "nvidia" ] || [ "$GPU_CLASS" = "apple" ] || [ "$GPU_CLASS" = "amd" ]; then
  RUNTIME="ollama"
fi
# Map VRAM → a model class. Apple/AMD: use a conservative mid class (unified memory varies).
if [ "$GPU_CLASS" = "nvidia" ]; then
  if   [ "$VRAM_GB" -ge 40 ] 2>/dev/null; then MODEL_CLASS="13B–34B instruct (e.g. llama3.1:8b → 34B class)"; MODEL_HINT="80GB cards serve 7B–13B at high throughput; 70B+ needs tensor-parallel across GPUs"
  elif [ "$VRAM_GB" -ge 20 ] 2>/dev/null; then MODEL_CLASS="7B–13B instruct (e.g. llama3.1:8b)"; MODEL_HINT="comfortable for an 8B model with room for context"
  elif [ "$VRAM_GB" -ge 10 ] 2>/dev/null; then MODEL_CLASS="7B–8B instruct, quantized (e.g. llama3.2:3b / llama3.1:8b-q4)"; MODEL_HINT="prefer a 3B–8B quantized model"
  elif [ "$VRAM_GB" -gt 0 ] 2>/dev/null;  then MODEL_CLASS="1B–3B instruct (e.g. llama3.2:1b / llama3.2:3b)"; MODEL_HINT="small VRAM — keep the model small or use echo"
  else MODEL_CLASS="7B–8B instruct (VRAM unknown — start small)"; fi
elif [ "$GPU_CLASS" = "apple" ] || [ "$GPU_CLASS" = "amd" ]; then
  MODEL_CLASS="3B–8B instruct (e.g. llama3.2:3b / llama3.1:8b)"; MODEL_HINT="run Ollama natively; size to your unified/GPU memory"
fi
ok "Runtime:  $RUNTIME"
ok "Model:    $MODEL_CLASS"
no "$MODEL_HINT"
no "You advertise a Basis tier (basis-small / basis-default / basis-large) while serving the real local model behind it."

# ── 6. Fetch the skill package ────────────────────────────────────────────────
fetch() { # fetch <url> <dest>
  if have curl; then curl -fsSL "$1" -o "$2";
  elif have wget; then wget -qO "$2" "$1";
  else return 9; fi
}
if [ "$DO_FETCH" = 1 ]; then
  hdr "6. Fetch the basis-worker skill"
  if ! have curl && ! have wget; then
    warn "Neither curl nor wget is available — cannot fetch. Install one, or copy the skill from the repo (.agents/skills/basis-worker)."
  else
    mkdir -p "$SKILL_DIR/scripts"
    SKILL_URL="$BASE_URL/api/skills/basis-worker"
    fetched=0
    if fetch "$SKILL_URL/SKILL.md" "$SKILL_DIR/SKILL.md"; then ok "SKILL.md → $SKILL_DIR/SKILL.md"; fetched=1; else warn "could not fetch SKILL.md"; fi
    if fetch "$SKILL_URL/manifest" "$SKILL_DIR/manifest.json"; then ok "manifest → $SKILL_DIR/manifest.json"; fi
    # Pull each script the download index lists (best-effort; the agent can also
    # copy them from the repo). The two markdown files are served raw; scripts are
    # mirrored under /skills/basis-worker/scripts/ for direct fetch.
    for s in setup-worker.mjs doctor.mjs publish-prices.mjs verify-worker.mjs healthcheck.mjs; do
      if fetch "$BASE_URL/skills/basis-worker/scripts/$s" "$SKILL_DIR/scripts/$s" 2>/dev/null; then
        no "scripts/$s"
      fi
    done
    if [ "$fetched" = 1 ]; then ok "Skill package ready in $SKILL_DIR"; fi
  fi
else
  hdr "6. Fetch"
  no "Skipped (--no-fetch). Detection + recommendation only."
fi

# ── 7. Next steps ─────────────────────────────────────────────────────────────
hdr "Next steps"
say "  1. Read the skill:        $SKILL_DIR/SKILL.md  (or $BASE_URL/api/skills/basis-worker/SKILL.md)"
say "  2. Confirm hardware:      node $SKILL_DIR/scripts/doctor.mjs"
say "  3. Register (PUBLIC wallet only — never a key/seed):"
say "       node $SKILL_DIR/scripts/setup-worker.mjs --register \\"
say "         --reward-wallet 0xYourPublicEvmAddress0000000000000000000000 \\"
say "         --worker-id my-gpu-1 --base $BASE_URL"
say "  4. Publish a signed price/model offer (the market):"
say "       node $SKILL_DIR/scripts/publish-prices.mjs --worker-id my-gpu-1 --prices ./prices.json --key-file ./worker-reward.key"
say "  5. Start the worker (standalone runtime/worker CLI):"
say "       cd runtime/worker && npm install && BASIS_WORKER_RUNTIME=$RUNTIME npm run dev"
say "  6. Check live posture:    curl -fsSL $BASE_URL/api/workers/onboarding/status"
say ""
say "  Honest state: worker registration + offer storage are live. PUBLIC worker"
say "  routing for /api/v1 is READY BUT OFF — a registered worker receives"
say "  production jobs only when routing is enabled AND a real verified worker is"
say "  online. On-chain settlement is keeper/operator-driven (not automatic)."
say "  Rewards are payment for completed verified work — never guaranteed, no yield."
say ""