install.sh

#!/usr/bin/env bash
# gpu-optimizer install script
# Sets up the vLLM prompt optimizer pipeline for Claude Code.
# Run from the repo root: ./install.sh
set -euo pipefail

REPO_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
HOOK_DEST="$HOME/.claude/hooks"
ASSET_DEST="$HOME/.claude/gpu-optimizer"
CONF_FILE="$HOME/.claude/gpu-optimizer.conf"
SETTINGS_FILE="$HOME/.claude/settings.json"

GREEN='\033[0;32m'; YELLOW='\033[1;33m'; CYAN='\033[0;36m'; NC='\033[0m'

info()    { echo -e "${CYAN}[gpu-optimizer]${NC} $*"; }
success() { echo -e "${GREEN}✓${NC} $*"; }
warn()    { echo -e "${YELLOW}⚠${NC}  $*"; }

# Fix 🔴: single-quote-escape a value for safe embedding in a sourced bash file.
# Replaces every ' with '\'' so the value is always treated as a literal string.
safe_val() { printf "'%s'" "${1//\'/\'\\\'\'}"; }

# Fix 🟡: validate a Kubernetes resource name — [a-z0-9][a-z0-9-]*, max 63 chars.
validate_k8s_name() {
  local name="$1" label="$2"
  if [[ ! "$name" =~ ^[a-z0-9][a-z0-9-]{0,62}$ ]]; then
    echo "Error: '$name' is not a valid Kubernetes $label name." \
         "Use lowercase letters, digits, and hyphens only." >&2
    exit 1
  fi
}

echo ""
echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
echo "  gpu-optimizer — Claude Code vLLM pipeline installer"
echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
echo ""

# ── check dependencies ────────────────────────────────────────────────────────
for dep in kubectl jq curl shasum; do
  if ! command -v "$dep" >/dev/null 2>&1; then
    echo "Error: '$dep' is required but not installed." >&2
    exit 1
  fi
done
success "Dependencies: kubectl, jq, curl, shasum — all found"

# ── collect config ────────────────────────────────────────────────────────────
echo ""
info "Configuration (press Enter to accept defaults)"
echo ""

read -rp "  vLLM model name          [qwen2.5-14b]: " VLLM_MODEL
VLLM_MODEL="${VLLM_MODEL:-qwen2.5-14b}"

read -rp "  Kubernetes namespace     [default]: " VLLM_NAMESPACE
VLLM_NAMESPACE="${VLLM_NAMESPACE:-default}"

read -rp "  Deployment name          [llm-inference]: " VLLM_DEPLOYMENT
VLLM_DEPLOYMENT="${VLLM_DEPLOYMENT:-llm-inference}"

read -rp "  Service name             [llm-inference]: " VLLM_SERVICE
VLLM_SERVICE="${VLLM_SERVICE:-llm-inference}"

read -rp "  PVC name                 [model-cache]: " VLLM_PVC
VLLM_PVC="${VLLM_PVC:-model-cache}"

read -rp "  Min prompt words to optimize [15]: " VLLM_MIN_WORDS
VLLM_MIN_WORDS="${VLLM_MIN_WORDS:-15}"

VLLM_ENDPOINT="http://localhost:8000"
VLLM_MANIFEST="$ASSET_DEST/llm-inference.yaml"

# Fix 🟡 + 🔴: validate k8s names before they reach the config file or any command string.
# This also closes the STOP_CMD single-quote injection: validated names can only be [a-z0-9-].
validate_k8s_name "$VLLM_NAMESPACE"  "namespace"
validate_k8s_name "$VLLM_DEPLOYMENT" "deployment"
validate_k8s_name "$VLLM_SERVICE"    "service"
validate_k8s_name "$VLLM_PVC"        "PVC"

# ── write config file ─────────────────────────────────────────────────────────
echo ""
info "Writing config to $CONF_FILE..."
# Fix 🔴: use safe_val() so user-supplied values are single-quote-escaped.
# A value like `x" && rm -rf ~` becomes `'x" && rm -rf ~'` — a literal string when sourced.
cat > "$CONF_FILE" <<EOF
# gpu-optimizer configuration — generated by install.sh
VLLM_ENDPOINT=$(safe_val "$VLLM_ENDPOINT")
VLLM_MODEL=$(safe_val "$VLLM_MODEL")
VLLM_NAMESPACE=$(safe_val "$VLLM_NAMESPACE")
VLLM_SERVICE=$(safe_val "$VLLM_SERVICE")
VLLM_DEPLOYMENT=$(safe_val "$VLLM_DEPLOYMENT")
VLLM_PVC=$(safe_val "$VLLM_PVC")
VLLM_MANIFEST=$(safe_val "$VLLM_MANIFEST")
VLLM_MIN_WORDS=$(safe_val "$VLLM_MIN_WORDS")
EOF
# Fix 🟠: restrict to owner read/write — conf contains infrastructure names.
chmod 600 "$CONF_FILE"
success "Config written (permissions: 600)"

# ── install hooks ─────────────────────────────────────────────────────────────
info "Installing hooks to $HOOK_DEST..."
mkdir -p "$HOOK_DEST"
for script in prompt-optimizer.sh fetch-compressor.sh gpu-scheduler.sh optimizer-status.sh llm-autoscale-down.sh; do
  cp "$REPO_DIR/hooks/$script" "$HOOK_DEST/$script"
  chmod +x "$HOOK_DEST/$script"
done
success "5 hooks installed"

# ── install k8s manifest ──────────────────────────────────────────────────────
info "Installing k8s manifest to $ASSET_DEST..."
mkdir -p "$ASSET_DEST"
cp "$REPO_DIR/k8s/llm-inference.yaml" "$ASSET_DEST/llm-inference.yaml"
success "Manifest installed"

# ── patch settings.json ───────────────────────────────────────────────────────
info "Patching $SETTINGS_FILE..."

# Ensure settings.json exists
[ -f "$SETTINGS_FILE" ] || echo '{}' > "$SETTINGS_FILE"

# Fix 🟠: back up before any mutation — settings.json controls all Claude Code hooks.
cp "$SETTINGS_FILE" "${SETTINGS_FILE}.bak"
info "Backup saved to ${SETTINGS_FILE}.bak"

HOOK_DIR="$HOOK_DEST"

# Build the hooks patch with jq — merges into existing hooks without clobbering other keys
STATS_RESET='{"session_prompts":0,"session_tokens_saved":0,"last_orig_tokens":0,"last_opt_tokens":0,"last_mode":"-"}'
STATS_FILE_PATH="$HOME/.claude/gpu-optimizer-stats.json"
SESSION_CMD="${HOOK_DIR}/gpu-scheduler.sh && printf '%s' '${STATS_RESET}' > ${STATS_FILE_PATH}"
STOP_CMD="pkill -f 'kubectl port-forward svc/${VLLM_SERVICE}' 2>/dev/null || true"

# Fix 🟠: write to a temp file then mv — mv on the same filesystem is atomic,
# so a crash or interrupt can never leave settings.json truncated/empty.
SETTINGS_TMP=$(mktemp "${SETTINGS_FILE}.XXXXXX")
PATCHED=$(jq \
  --arg opt     "$HOOK_DIR/prompt-optimizer.sh" \
  --arg comp    "$HOOK_DIR/fetch-compressor.sh" \
  --arg stat    "$HOOK_DIR/optimizer-status.sh" \
  --arg session "$SESSION_CMD" \
  --arg stop    "$STOP_CMD" \
  '
  .hooks.UserPromptSubmit = [
    {
      hooks: [{
        type: "command",
        command: $opt,
        timeout: 20,
        statusMessage: "Optimizing prompt..."
      }]
    }
  ] |
  .hooks.PostToolUse = (
    (.hooks.PostToolUse // []) |
    map(select(.matcher != "WebFetch")) +
    [{
      matcher: "WebFetch",
      hooks: [{
        type: "command",
        command: $comp,
        timeout: 20,
        statusMessage: "Compressing fetch..."
      }]
    }]
  ) |
  .hooks.SessionStart = [
    {hooks: [{type: "command", command: $session, async: true}]}
  ] |
  .hooks.Stop = [
    {hooks: [{type: "command", command: $stop, async: true}]}
  ] |
  .statusLine = {
    type: "command",
    command: $stat,
    refreshInterval: 10
  }
  ' "$SETTINGS_FILE") && { echo "$PATCHED" > "$SETTINGS_TMP" && mv "$SETTINGS_TMP" "$SETTINGS_FILE"; }

success "settings.json patched"

# ── set up cron ───────────────────────────────────────────────────────────────
info "Setting up auto-scale-down cron (every 15 min)..."
CRON_CMD="*/15 * * * * $HOOK_DIR/llm-autoscale-down.sh >> $HOME/.claude/gpu-optimizer-autoscale.log 2>&1"

# Remove any existing gpu-optimizer cron line, add the new one.
# `crontab -l` exits non-zero when no crontab exists; `|| true` prevents that
# from killing the subshell under set -euo pipefail before echo runs.
EXISTING_CRON=$(crontab -l 2>/dev/null || true)
( echo "$EXISTING_CRON" | grep -v "llm-autoscale-down\|gpu-optimizer-autoscale" ; echo "$CRON_CMD" ) | crontab -
success "Cron job installed"

# ── done ──────────────────────────────────────────────────────────────────────
echo ""
echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
echo -e "${GREEN}  Installation complete!${NC}"
echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
echo ""
echo "  Next steps:"
echo "  1. Apply the k8s manifest:"
echo "     kubectl apply -f $ASSET_DEST/llm-inference.yaml"
echo ""
echo "  2. Restart Claude Code — hooks are active on next session."
echo ""
echo "  3. Optionally edit $CONF_FILE to tune settings."
echo ""
warn "  The k8s manifest targets Qwen2.5-14B-AWQ on nvidia-l4 Spot."
warn "  Edit $ASSET_DEST/llm-inference.yaml to change the model or GPU type."
echo ""