#!/usr/bin/env bash
set -Eeuo pipefail

config_dir="${XDG_CONFIG_HOME:-$HOME/.config}/hermes-stack-maintenance"
CONFIG_FILE="${HERMES_LLAMA_CONFIG:-$config_dir/llama.env}"
LLAMA_ARGS=()
# shellcheck disable=SC1090
source "$CONFIG_FILE"

mode=run
profile_override=""
while (($#)); do
  case "$1" in
    --check) mode=check ;;
    --print) mode=print ;;
    --profile)
      [[ $# -ge 2 ]] || exit 2
      profile_override="$2"
      shift
      ;;
    *)
      printf 'Aufruf: %s [--check|--print] [--profile NAME]\n' "$0" >&2
      exit 2
      ;;
  esac
  shift
done

if [[ -n "${LLAMA_PROFILE:-}" && -r "$config_dir/llama-profile.env" ]]; then
  # shellcheck disable=SC1090,SC1091
  source "$config_dir/llama-profile.env"
fi
[[ -z "$profile_override" ]] || LLAMA_PROFILE="$profile_override"

if [[ -z "${LLAMA_SERVER_BIN:-}" ]]; then
  LLAMA_SERVER_BIN="$HOME/.local/opt/llama.cpp/current/bin/llama-server"
  if [[ ! -x "$LLAMA_SERVER_BIN" ]]; then
    LLAMA_SERVER_BIN="$(command -v llama-server || true)"
  fi
fi

[[ -x "$LLAMA_SERVER_BIN" ]] || {
  printf 'llama-server nicht ausführbar: %s\n' "$LLAMA_SERVER_BIN" >&2
  exit 1
}

model_args=()
if [[ -n "${LLAMA_MODEL:-}" && -n "${LLAMA_HF_REPO:-}" ]]; then
  printf 'Nur LLAMA_MODEL ODER LLAMA_HF_REPO setzen.\n' >&2
  exit 1
elif [[ -n "${LLAMA_MODEL:-}" ]]; then
  [[ -r "$LLAMA_MODEL" ]] || {
    printf 'Modell nicht lesbar.\n' >&2
    exit 1
  }
  model_args=(--model "$LLAMA_MODEL")
elif [[ -n "${LLAMA_HF_REPO:-}" ]]; then
  model_args=(--hf-repo "$LLAMA_HF_REPO")
else
  printf 'LLAMA_MODEL oder LLAMA_HF_REPO fehlt.\n' >&2
  exit 1
fi

if [[ -n "${LLAMA_MMPROJ:-}" ]]; then
  [[ -r "$LLAMA_MMPROJ" ]] || {
    printf 'Projektor nicht lesbar.\n' >&2
    exit 1
  }
  model_args+=(--mmproj "$LLAMA_MMPROJ")
fi
[[ -z "${LLAMA_ALIAS:-}" ]] || model_args+=(--alias "$LLAMA_ALIAS")

# Server defaults are not a wall-clock limit; the independent slot watchdog is.
[[ "${LLAMA_MAX_OUTPUT_TOKENS:-2048}" =~ ^[1-9][0-9]*$ ]] || exit 2

profile_args=()
if [[ -n "${LLAMA_PROFILE:-}" ]]; then
  reasoning_budget=512
  reasoning_preserve=--reasoning-preserve
  case "$LLAMA_PROFILE" in
    desktop-64k) ctx=65536; margin=4096; reasoning_budget=128; reasoning_preserve=--no-reasoning-preserve ;;
    baseline) ctx=65536; margin=1536 ;;
    vram-64k) ctx=65536; margin=1280 ;;
    balanced) ctx=49152; margin=1536 ;;
    vram-plus) ctx=49152; margin=1280 ;;
    compact) ctx=32768; margin=1536 ;;
    *)
      printf 'Unbekanntes Profil: %s\n' "$LLAMA_PROFILE" >&2
      exit 2
      ;;
  esac

  profile_args=(
    --ctx-size "$ctx" --parallel 1 --jinja
    --reasoning auto --reasoning-budget "$reasoning_budget" "$reasoning_preserve"
    --flash-attn auto --threads 6 --threads-batch 16
    --n-gpu-layers auto --fit on --fit-target "$margin"
    --cache-type-k q8_0 --cache-type-v q8_0
    --sleep-idle-seconds 300 --image-min-tokens 1024
    --cors-origins localhost --no-ui
  )

  if ((${#LLAMA_ARGS[@]})); then
    printf 'Bei Profilbetrieb LLAMA_ARGS leeren; eigene vollständige Args nur ohne LLAMA_PROFILE.\n' >&2
    exit 2
  fi
fi

args=(
  "${model_args[@]}"
  --host "${LLAMA_HOST:-127.0.0.1}"
  --port "${LLAMA_PORT:-8080}"
  --n-predict "${LLAMA_MAX_OUTPUT_TOKENS:-2048}"
  "${profile_args[@]}"
  "${LLAMA_ARGS[@]}"
)

if [[ "$mode" == print ]]; then
  printf '%q ' "$LLAMA_SERVER_BIN" "${args[@]}"
  printf '\n'
  exit 0
fi

# Nur Optionsnamen prüfen: kein Modellstart oder Download.
help_text="$(timeout 30 "$LLAMA_SERVER_BIN" --help 2>&1)" || {
  printf 'llama-server --help fehlgeschlagen.\n' >&2
  exit 1
}

for arg in "${args[@]}"; do
  [[ "$arg" == --* ]] || continue
  option="${arg%%=*}"
  if ! grep -Eq -- "(^|[[:space:],])${option}([[:space:],=]|$)" <<<"$help_text"; then
    printf 'Die installierte llama.cpp-Version kennt %s nicht. Kein automatisches Weglassen.\n' "$option" >&2
    exit 1
  fi
done

if [[ "$mode" == check ]]; then
  printf 'Konfiguration und Optionsnamen OK; GPU/Modell/Optionswerte noch nicht laufzeitgetestet.\n'
  exit 0
fi

exec "$LLAMA_SERVER_BIN" "${args[@]}"