#!/usr/bin/env bash
# Fable scores one app's readability three times, blind to which agent or
# app it is (only the language is named); the per-field median is kept.
#   builders/run-readability.sh <slug>
set -euo pipefail
slug=$1
repo=$(cd "$(dirname "$0")/.." && pwd)
cd "$repo"
language=$(cd harness && go run ./cmd/language "$slug")
work=$(mktemp -d -p "${TMPDIR:-$HOME/.cache}")
trap 'rm -rf "$work"' EXIT
{
  sed "s/{LANGUAGE}/$language/g" builders/prompts/readability.md
  echo
  bakeoff source --slug "$slug" --dir "${APP_DIR:-apps/$slug}" | sed -e "s/todo-$slug/todo-app/g" -e "s|apps/$slug|apps/app|g"
} > "$work/prompt.md"
replies=()
for i in 1 2 3; do
  claude -p --model claude-fable-5-1 --output-format json --tools "" \
    --mcp-config '{"mcpServers":{}}' --strict-mcp-config --no-session-persistence \
    < "$work/prompt.md" > "builders/logs/$slug.readability$i.json"
  python3 -c 'import json,sys; print(json.load(open(sys.argv[1]))["result"])' "builders/logs/$slug.readability$i.json" > "$work/reply$i"
  replies+=("$work/reply$i")
done
bakeoff readability --slug "$slug" --replies "$(IFS=,; echo "${replies[*]}")"