#!/bin/bash
# xwcode 功能验证测试套件
# 用法:
#   XWCODE_BIN=./target/release/xwcode bash scripts/test_suite.sh
#   远程: ssh km "XWCODE_BIN=~/bin/xwcode bash ~/xwcode/scripts/test_suite.sh"
set -euo pipefail

BIN="${XWCODE_BIN:-./target/release/xwcode}"
TIMEOUT=60
PASS=0; FAIL=0

# 日志归档目录（每次测试独立保存）
RUN_ID="test_$(date +%Y%m%d_%H%M%S)"
LOG_DIR="${XWCODE_LOG_DIR:-./.xwcode/test_runs/$RUN_ID}"
EVENT_SRC="${XWCODE_EVENT_DIR:-./.xwcode/events}"

green() { echo -e "\033[32m$1\033[0m"; }
red()   { echo -e "\033[31m$1\033[0m"; }
dim()   { echo -e "\033[2m$1\033[0m"; }

# oneshot 单轮测试（带日志采集）
test_oneshot() {
    local id="$1" desc="$2" prompt="$3" grep_for="$4" grep_not="${5:-}"
    echo -n "  [$id] $desc ... "
    local start=$(date +%s)
    local output
    output=$(timeout $TIMEOUT "$BIN" -p "$prompt" --max-steps 5 --no-tui 2>&1) || true
    local elapsed=$(($(date +%s) - start))

    # 保存本轮完整输出
    echo "$output" > "$LOG_DIR/${id}_output.txt"

    local result="PASS"
    if [ -n "$grep_for" ] && ! echo "$output" | grep "$grep_for" >/dev/null; then
        result="FAIL: expected '$grep_for' not found"
    fi
    if [ -n "$grep_not" ] && echo "$output" | grep "$grep_not" >/dev/null; then
        result="FAIL: unwanted '$grep_not' found"
    fi

    if [ "$result" = "PASS" ]; then
        green "PASS (${elapsed}s)"; PASS=$((PASS+1))
    else
        red "$result (${elapsed}s)"; FAIL=$((FAIL+1))
        dim "  prompt: $prompt"
        dim "  output: $(echo "$output" | grep -E '\[tool\]|tool err|tool ok|Error|未知工具|权限' | head -5)"
    fi
}

# ── 初始化 ──────────────────────────────────────

mkdir -p "$LOG_DIR"
echo "=== xwcode 功能验证 ($(date '+%Y-%m-%d %H:%M:%S')) ===" | tee "$LOG_DIR/summary.txt"
echo "Binary: $BIN" | tee -a "$LOG_DIR/summary.txt"
echo "Log:    $LOG_DIR" | tee -a "$LOG_DIR/summary.txt"
echo "" | tee -a "$LOG_DIR/summary.txt"

# ── P0: 基础交互 ──
echo "── P0: 基础交互 ──" | tee -a "$LOG_DIR/summary.txt"
test_oneshot "T01" "问候不调工具" \
    "你好，简单打个招呼" \
    "" \
    "\[tool\]"

test_oneshot "T02" "简单问答不调工具" \
    "Rust 中 Vec 和数组有什么区别？用文字简要回答即可" \
    "" \
    "\[tool\]"

test_oneshot "T03" "代码阅读可用 tree_view" \
    "tree_view ." \
    "tool ok"

test_oneshot "T04" "代码阅读可用 read_file" \
    "读取 Cargo.toml 文件内容" \
    "tool ok"

test_oneshot "T05" "代码搜索可用 grep_search" \
    "搜索项目中 main 函数的定义" \
    "tool ok"

# ── P1: 无回归 ──
echo "── P1: 无回归 ──" | tee -a "$LOG_DIR/summary.txt"
test_oneshot "T11" "tree_view 不报未知工具" \
    "调用 tree_view . 查看目录" \
    "" \
    "未知工具.*tree_view"

test_oneshot "T12" "不报权限拒绝" \
    "记住测试项：xwcode 冒烟测试通过" \
    "" \
    "权限拒绝\|DangerFullAccess"

# ── P2: 分析连续性 ──
echo "── P2: 分析连续性 ──" | tee -a "$LOG_DIR/summary.txt"
test_oneshot "T21" "分析任务允许多次调用" \
    "分析 Cargo.toml 和 Cargo.lock 两个文件，总结项目依赖" \
    "tool ok" \
    ""

# ── 采集日志 ────────────────────────────────────

echo "" | tee -a "$LOG_DIR/summary.txt"
echo "── 采集 EventLog ──" | tee -a "$LOG_DIR/summary.txt"

# 采集最新的事件日志
if [ -d "$EVENT_SRC" ]; then
    # model_calls — LLM 调用记录（prompt/response/tool_calls/token）
    if [ -f "$EVENT_SRC/model_calls.jsonl" ]; then
        tail -50 "$EVENT_SRC/model_calls.jsonl" > "$LOG_DIR/model_calls.jsonl" 2>/dev/null || true
        echo "  model_calls: $(wc -l < "$LOG_DIR/model_calls.jsonl") lines" | tee -a "$LOG_DIR/summary.txt"
    fi
    # tool_outputs — 工具调用输入输出
    if [ -f "$EVENT_SRC/tool_outputs.jsonl" ]; then
        tail -50 "$EVENT_SRC/tool_outputs.jsonl" > "$LOG_DIR/tool_outputs.jsonl" 2>/dev/null || true
        echo "  tool_outputs: $(wc -l < "$LOG_DIR/tool_outputs.jsonl") lines" | tee -a "$LOG_DIR/summary.txt"
    fi
    # phase_timing — 阶段耗时
    if [ -f "$EVENT_SRC/phase_timing.jsonl" ]; then
        tail -20 "$EVENT_SRC/phase_timing.jsonl" > "$LOG_DIR/phase_timing.jsonl" 2>/dev/null || true
        echo "  phase_timing: $(wc -l < "$LOG_DIR/phase_timing.jsonl") lines" | tee -a "$LOG_DIR/summary.txt"
    fi
fi

# ── 结果 ────────────────────────────────────────

echo "" | tee -a "$LOG_DIR/summary.txt"
echo "=== 结果: $PASS/8 pass ===" | tee -a "$LOG_DIR/summary.txt"
[ $FAIL -eq 0 ] && green "全部通过 ✅" || red "有 $FAIL 项失败 ❌"

# 写入测试报告
cat > "$LOG_DIR/report.json" << EOF
{
  "run_id": "$RUN_ID",
  "timestamp": "$(date -Iseconds)",
  "binary": "$BIN",
  "passed": $PASS,
  "failed": $FAIL,
  "total": 8,
  "log_dir": "$LOG_DIR"
}
EOF

echo "报告: $LOG_DIR/report.json"
exit $FAIL