#!/usr/bin/env bash
# Gets the repository ready to run: submodules, patches, binaries and models.
#
# It is idempotent: it can be run again safely. And it does not copy models
# (they are more than 5 GB); it links them from wherever they already are.
set -euo pipefail

cd "$(dirname "$0")/.."
ROOT="$PWD"

# Repositories already cloned on the system, so nothing is downloaded or
# rebuilt again. ASIST_ESPEJO is the old name of ASIST_MIRROR.
MIRROR="${ASIST_MIRROR:-${ASIST_ESPEJO:-$HOME/GIT-MIRRO}}"
HF="${ASIST_HF:-$HOME/HF}"

blue()   { printf '\033[36m%s\033[0m\n' "$*"; }
green()  { printf '\033[32m%s\033[0m\n' "$*"; }
warn()   { printf '\033[33m%s\033[0m\n' "$*" >&2; }
bad()    { printf '\033[31m%s\033[0m\n' "$*" >&2; }

link() { # source target description
  local src="$1" dst="$2" what="$3"
  if [ ! -e "$src" ]; then
    warn "  $what not found: $src"
    return 1
  fi
  mkdir -p "$(dirname "$dst")"
  if [ -L "$dst" ] || [ -e "$dst" ]; then
    rm -rf "$dst"
  fi
  ln -s "$src" "$dst"
  green "  $what -> $(basename "$src")"
}

# ---------------------------------------------------------------------------
blue "1/5  Submodules"
# If the repository is cloned next door, it is used as a reference: llama.cpp
# is 406 MB of objects that do not need downloading again.
for name in canary-rs qwentts.cpp llama.cpp; do
  path="vendor/$name"
  if [ -f "$path/.git" ] || [ -d "$path/.git" ]; then
    continue
  fi
  if [ -d "$MIRROR/$name/.git" ]; then
    git -c protocol.file.allow=always submodule update --init \
        --reference "$MIRROR/$name" "$path"
  else
    git submodule update --init "$path"
  fi
done
git submodule status | sed 's/^/  /'

# ---------------------------------------------------------------------------
blue "2/5  Patches and loose files"
# The patches hold the local changes on each repository: without them the ASR
# behaves differently from the one that was measured.
for name in canary-rs qwentts.cpp llama.cpp; do
  patch="$ROOT/vendor/patches/$name.patch"
  [ -s "$patch" ] || continue
  (
    cd "vendor/$name"
    if git apply --check "$patch" 2>/dev/null; then
      git apply "$patch"
      green "  $name: patch applied"
    elif git apply --reverse --check "$patch" 2>/dev/null; then
      green "  $name: patch already applied"
    else
      bad "  $name: the patch does not apply; review it by hand"
    fi
  )
done
# Files that only existed in the local working tree and a patch cannot carry.
# canary-rs does not build without bench_live.rs: its Cargo.toml declares it.
if [ -d vendor/extra ]; then
  for name in canary-rs qwentts.cpp; do
    [ -d "vendor/extra/$name" ] || continue
    cp -rn "vendor/extra/$name/." "vendor/$name/" 2>/dev/null || true
  done
  green "  loose files copied"
fi

# ---------------------------------------------------------------------------
blue "3/5  C++ engines"
link_or_build() { # name build_path build_command...
  local name="$1" build="$2"; shift 2
  if [ -e "vendor/$name/$build" ]; then
    green "  $name is already built"
    return
  fi
  if [ -e "$MIRROR/$name/$build" ]; then
    link "$MIRROR/$name/$(dirname "$build")" \
         "$ROOT/vendor/$name/$(dirname "$build")" "$name build"
    return
  fi
  warn "  $name not built. Building (this takes a while)…"
  ( cd "vendor/$name" && "$@" )
}
link_or_build llama.cpp build/bin/llama-server \
  bash -c 'cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON && cmake --build build -j --target llama-server'
link_or_build qwentts.cpp build/tts-server \
  bash -c './buildcuda.sh'

# ---------------------------------------------------------------------------
blue "4/5  Models"
# Linked, not copied: together they are more than 5 GB.
link "$HF/Qwen/Qwen3.5-2B.Q5_K_M.gguf" "$ROOT/models/Qwen3.5-2B.Q5_K_M.gguf" "LLM model" || true
link "$HF/Qwen/mmproj-BF16.gguf"       "$ROOT/models/mmproj-BF16.gguf"       "multimodal projector" || true
for m in qwen-talker-1.7b-base-Q8_0.gguf qwen-tokenizer-12hz-Q8_0.gguf; do
  link "$MIRROR/qwentts.cpp/models/$m" "$ROOT/models/$m" "$m" || true
done
if [ ! -d vendor/canary-rs/models/canary-180m-flash-onnx ] \
   && [ -d "$MIRROR/canary-rs/models/canary-180m-flash-onnx" ]; then
  link "$MIRROR/canary-rs/models/canary-180m-flash-onnx" \
       "$ROOT/vendor/canary-rs/models/canary-180m-flash-onnx" "Canary model" || true
fi

# ---------------------------------------------------------------------------
blue "5/5  Reference voice"
# The assistant speaks with a cloned voice. If none is prepared, say how to make
# one instead of failing: everything else already works with a model voice.
if [ -f assets/voices/asistente.spk ]; then
  green "  there is already a voice in assets/voices/"
else
  warn "  no cloned voice. Prepare one with:"
  warn "      scripts/clone-voice.sh <recording.wav> <transcript.txt>"
  warn "  or remove the [tts.reference] section from config/asistente.toml to"
  warn "  use one of the model's own voices."
fi

echo
blue "Done. Check with:  cargo run --release -- check"