@bolloon/bolloon-agent 0.1.33 → 0.1.35

package/scripts/auto-evolve-loop.ts

@@ -0,0 +1,376 @@
+#!/usr/bin/env tsx
+/**
+ * auto-evolve-loop.ts — 阶段 D 主循环 (重写版)
+ *
+ * 流程 (per iteration):
+ *   1. 跑 vitest, 拿 fail 信息
+ *   2. 把 fail + 当前 src/test 源码喂给 LLM (走 PiAIModel / MiniMax)
+ *   3. 写 diff 到 staging/auto-evolve/iter-<N>/
+ *   4. 跑 detect-schema-changes + diff-reviewer
+ *   5a. reviewer PASS → git apply --recount --whitespace=fix + commit
+ *   5b. reviewer FAIL → log, 累计连续失败
+ *   6. 连续 3 次 FAIL → 自动回滚 baseline + 通知
+ *
+ * 退出条件 (任一即停):
+ *   - 全部测试 PASS (loop 成功)
+ *   - 连续 3 次 FAIL (loop 认输)
+ *   - max-iter 达到 (默认 10)
+ *   - 人类 SIGINT (Ctrl-C)
+ *
+ * 用法:
+ *   tsx scripts/auto-evolve-loop.ts
+ *   tsx scripts/auto-evolve-loop.ts --max-iter 5
+ *   tsx scripts/auto-evolve-loop.ts --target src/utils/foo.ts
+ */
+
+import { execFile } from 'child_process';
+import * as fs from 'fs/promises';
+import { promisify } from 'util';
+import * as path from 'path';
+
+const pExec = promisify(execFile);
+const REPO = process.cwd();
+
+interface VitestResult {
+  failed: number;
+  passed: number;
+  totalFiles: string[];
+  failingTests: { file: string; name: string; message: string }[];
+}
+
+/**
+ * 跑全量 vitest, 在 parseVitestJson 里 filter
+ * (vitest 不接受 positional path filter, 会被忽略)
+ */
+async function runVitest(targetFile?: string): Promise<VitestResult> {
+  const args = ['vitest', 'run', '--reporter=json', '--no-color'];
+  try {
+    const { stdout } = await pExec('npx', args, { cwd: REPO, maxBuffer: 50 * 1024 * 1024 });
+    const json = JSON.parse(stdout);
+    return parseVitestJson(json, targetFile);
+  } catch (err: any) {
+    if (err.stdout) {
+      try {
+        const json = JSON.parse(err.stdout);
+        return parseVitestJson(json, targetFile);
+      } catch {
+        // 解析失败
+      }
+    }
+    return { failed: 1, passed: 0, totalFiles: [], failingTests: [{ file: 'unknown', name: 'vitest crashed', message: err.message?.slice(0, 500) || '' }] };
+  }
+}
+
+function parseVitestJson(j: any, targetFile?: string): VitestResult {
+  const failingTests: VitestResult['failingTests'] = [];
+  const totalFiles: string[] = [];
+  let failed = 0;
+  let passed = 0;
+  for (const f of j.testResults || []) {
+    // targetFile 过滤: 只看匹配的文件
+    if (targetFile && !f.name?.includes(targetFile.split('/').pop() || '')) {
+      continue;
+    }
+    totalFiles.push(f.name || '');
+    failed += f.assertionResults?.filter((a: any) => a.status === 'failed').length || 0;
+    passed += f.assertionResults?.filter((a: any) => a.status === 'passed').length || 0;
+    for (const a of f.assertionResults || []) {
+      if (a.status === 'failed') {
+        failingTests.push({
+          file: f.name,
+          name: a.fullName || a.title,
+          message: (a.failureMessages || []).join('\n').slice(0, 1500),
+        });
+      }
+    }
+  }
+  return { failed, passed, totalFiles, failingTests };
+}
+
+/**
+ * 调 LLM — 走 bolloon 自带 PiAIModel (MiniMax, 读 .env 的 MINIMAX_API_KEY)
+ * 无 key 或 SDK 不可用都返回空字符串 (loop 认作 "无 diff" 重试)
+ */
+async function callLLM(prompt: string): Promise<string> {
+  try {
+    const pi = await import('../src/llm/pi-ai.js' as any).catch(() => import('../src/llm/pi-ai.js'));
+    const client = pi.initMinimax();
+    const text = await client.generateText({
+      messages: [{ role: 'user', content: prompt }],
+      maxTokens: 4096,
+      temperature: 0.2,
+    });
+    await bumpCallCount();
+    return text || '';
+  } catch (e: any) {
+    console.warn(`[loop] PiAIModel 不可用: ${e.message?.slice(0, 100)}`);
+    return '';
+  }
+}
+
+async function bumpCallCount(): Promise<void> {
+  const f = path.join(REPO, '.auto-evolve-calls');
+  const today = new Date().toISOString().slice(0, 10);
+  let count = 0;
+  try {
+    const lines = (await fs.readFile(f, 'utf-8')).trim().split('\n').filter(Boolean);
+    const lastLine = lines[lines.length - 1];
+    if (lastLine?.startsWith(today)) count = parseInt(lastLine.split(':')[1] || '0', 10);
+  } catch { /* no file yet */ }
+  await fs.writeFile(f, `${today}:${count + 1}\n`, 'utf-8');
+}
+
+const FIX_PROMPT = `你是一个谨慎的代码修复助手. 你的工作是修复失败的测试, **不**做无关改动.
+
+约束 (违反任一即重做):
+  1. 改动最小: 只动让测试通过必需的部分
+  2. 不改测试本身 (除非 test 自身有 bug)
+  3. 不引入新依赖
+  4. 不删注释 / 不改注释
+  5. 不动 schema (interface/type) 除非 fail 信息明确要求
+  6. 不用 any / unknown / @ts-ignore 偷懒
+  7. 不动 package.json / tsconfig.json
+
+输出格式 (严格):
+  - 第一个字符必须是 \`\`\`diff
+  - 中间是 unified diff (git format-patch 风格, --- a/path +++ b/path)
+  - 最后一个字符必须是 \`\`\`
+  - 不要在 diff 块外输出任何文字
+  - 不要使用 thinking 块 (下游会丢失)
+  - **改动必须严格匹配当前源码 (行数, 缩进)**
+
+FAIL 信息:
+{{FAIL}}
+
+相关源码 (请严格匹配行数和缩进):
+{{SOURCE}}
+
+请**只**输出 diff 块:`;
+
+async function getSourceContext(file: string): Promise<string> {
+  try {
+    const content = await fs.readFile(file, 'utf-8');
+    return content.slice(0, 8000);
+  } catch {
+    return '';
+  }
+}
+
+/**
+ * 从 test 文件的 import 语句里提取 src 路径, 读出来
+ * 关键: 让 LLM 看到真实源文件, 否则它只能猜行数, diff apply 会失败
+ */
+async function getImportedSource(testFile: string): Promise<string> {
+  try {
+    const testContent = await fs.readFile(testFile, 'utf-8');
+    const dir = path.dirname(testFile);
+    const imports = [...testContent.matchAll(/from\s+['"]([^'"]+)['"]/g)].map((m) => m[1]);
+    const out: string[] = [];
+    for (const imp of imports) {
+      if (!imp.startsWith('.')) continue;
+      let resolved = path.resolve(dir, imp);
+      if (resolved.endsWith('.js')) {
+        // 试 .ts 替换
+        const tsTry = resolved.replace(/\.js$/, '.ts');
+        if (await fs.access(tsTry).then(() => true).catch(() => false)) {
+          resolved = tsTry;
+        }
+      } else if (!resolved.endsWith('.ts')) {
+        if (await fs.access(resolved + '.ts').then(() => true).catch(() => false)) {
+          resolved += '.ts';
+        }
+      }
+      try {
+        const content = await fs.readFile(resolved, 'utf-8');
+        out.push(`--- ${resolved} ---\n${content.slice(0, 4000)}`);
+      } catch { /* skip */ }
+    }
+    return out.join('\n\n');
+  } catch {
+    return '';
+  }
+}
+
+function extractDiff(llmOutput: string): string | null {
+  // 找 ```diff ... ``` 块
+  const m = /```diff\s*([\s\S]*?)```/.exec(llmOutput);
+  if (m) return m[1].trim();
+  return null;
+}
+
+async function writePatch(iter: number, patchContent: string): Promise<string> {
+  const id = `iter-${String(iter).padStart(3, '0')}`;
+  const dir = path.join(REPO, 'staging', 'auto-evolve', id);
+  await fs.mkdir(dir, { recursive: true });
+  await fs.writeFile(path.join(dir, `${id}.patch`), patchContent, 'utf-8');
+  await fs.writeFile(path.join(dir, '.patch-id'), id, 'utf-8');
+  return id;
+}
+
+async function runReviewer(patchId: string): Promise<{ verdict: 'PASS' | 'FAIL'; concerns: string[] }> {
+  try {
+    await pExec('bash', [path.join(REPO, 'scripts/detect-schema-changes.sh'), patchId], { cwd: REPO });
+    await pExec('npx', ['tsx', 'scripts/diff-reviewer.ts', patchId, '--model', 'claude-sonnet-4-6'], { cwd: REPO });
+    const verdictFile = path.join(REPO, 'staging', 'auto-evolve', patchId, '.review-verdict');
+    const json = JSON.parse(await fs.readFile(verdictFile, 'utf-8'));
+    return { verdict: json.verdict, concerns: json.concerns || [] };
+  } catch (err: any) {
+    return { verdict: 'FAIL', concerns: [err.message?.slice(0, 200) || 'reviewer crashed'] };
+  }
+}
+
+async function commitPatch(patchId: string): Promise<boolean> {
+  try {
+    const patchFile = path.join(REPO, 'staging', 'auto-evolve', patchId, `${patchId}.patch`);
+
+    // patch 末尾补 newline (LLM 输出可能无)
+    let patchContent = await fs.readFile(patchFile, 'utf-8');
+    if (!patchContent.endsWith('\n')) patchContent += '\n';
+    await fs.writeFile(patchFile, patchContent, 'utf-8');
+
+    // --whitespace=fix 容错, --recount 重算行号
+    await pExec('git', ['apply', '--recount', '--whitespace=fix', patchFile], { cwd: REPO });
+    const { stdout } = await pExec('git', ['status', '--porcelain'], { cwd: REPO });
+    const files = stdout.trim().split('\n').filter(Boolean).map((l) => l.split(/\s+/).slice(1).join(' '));
+    for (const f of files) {
+      try {
+        await pExec('git', ['add', f], { cwd: REPO });
+      } catch { /* binary or removed */ }
+    }
+    await pExec('git', ['commit', '-m', `auto-evolve: ${patchId} (LLM 修复)`], { cwd: REPO });
+    return true;
+  } catch (err: any) {
+    console.error(`[loop] commit 失败: ${err.message?.slice(0, 200)}`);
+    return false;
+  }
+}
+
+async function rollback(baseline: string): Promise<void> {
+  console.log(`[loop] ⚠️  自动回滚到 ${baseline}`);
+  await pExec('git', ['reset', '--hard', baseline], { cwd: REPO });
+}
+
+async function notify(msg: string): Promise<void> {
+  const f = path.join(REPO, '.auto-evolve-notify');
+  const line = `[${new Date().toISOString()}] ${msg}\n`;
+  await fs.appendFile(f, line, 'utf-8');
+  console.log(`[notify] ${msg}`);
+}
+
+async function main() {
+  const args = process.argv.slice(2);
+  let maxIter = 10;
+  let targetFile: string | undefined;
+  for (let i = 0; i < args.length; i++) {
+    if (args[i] === '--max-iter' && args[i + 1]) maxIter = parseInt(args[++i], 10);
+    if (args[i] === '--target' && args[i + 1]) targetFile = args[++i];
+  }
+
+  // 必须先有 baseline
+  if (!(await fs.stat('.last-auto-evolve-baseline').catch(() => null))) {
+    console.log('[loop] 没 baseline, 先打一个');
+    await pExec('bash', ['scripts/auto-evolve-snapshot.sh', 'snapshot'], { cwd: REPO });
+  }
+  const baseline = (await fs.readFile('.last-auto-evolve-baseline', 'utf-8')).trim();
+  console.log(`[loop] baseline: ${baseline}, max-iter: ${maxIter}`);
+
+  let consecutiveFails = 0;
+  for (let iter = 1; iter <= maxIter; iter++) {
+    console.log(`\n========== iter ${iter}/${maxIter} ==========`);
+
+    // 1. 跑 vitest
+    const result = await runVitest(targetFile);
+    console.log(`[vitest] passed=${result.passed} failed=${result.failed}`);
+
+    if (result.failed === 0) {
+      console.log('[loop] ✅ 全部测试通过, 退出');
+      return;
+    }
+
+    // 2. 让 LLM 修
+    const failSummary = result.failingTests
+      .slice(0, 3)
+      .map((f) => `FILE: ${f.file}\nTEST: ${f.name}\nERROR: ${f.message}`)
+      .join('\n---\n');
+    const firstTest = result.failingTests[0]?.file || '';
+    const testCtx = await getSourceContext(firstTest);
+    const srcCtx = await getImportedSource(firstTest);
+    const sourceCtx = `=== TEST FILE ===\n${testCtx}\n\n=== SRC FILE (imported) ===\n${srcCtx}`;
+    const prompt = FIX_PROMPT
+      .replace('{{FAIL}}', failSummary)
+      .replace('{{SOURCE}}', sourceCtx);
+
+    console.log('[loop] 调 LLM 修...');
+    const llmOut = await callLLM(prompt);
+    const diff = extractDiff(llmOut);
+    if (!diff) {
+      console.log('[loop] LLM 没返回有效 diff');
+      console.log('--- LLM 原始输出 (前 800) ---');
+      console.log(llmOut.slice(0, 800));
+      console.log('---');
+      consecutiveFails++;
+      if (consecutiveFails >= 3) {
+        await notify(`连续 3 次 LLM 无 diff, 自动回滚 ${baseline}`);
+        await rollback(baseline);
+        return;
+      }
+      continue;
+    }
+
+    // 3. 写 staging
+    const patchId = await writePatch(iter, diff);
+    console.log(`[loop] 写到 ${patchId}`);
+
+    // 4. 跑 reviewer
+    const review = await runReviewer(patchId);
+    console.log(`[reviewer] verdict=${review.verdict} concerns=${review.concerns.length}`);
+    if (review.concerns.length > 0) {
+      for (const c of review.concerns) console.log(`  - ${c}`);
+    }
+
+    if (review.verdict === 'FAIL') {
+      consecutiveFails++;
+      if (consecutiveFails >= 3) {
+        await notify(`连续 3 次 reviewer FAIL, 自动回滚 ${baseline}`);
+        await rollback(baseline);
+        return;
+      }
+      continue;
+    }
+
+    // 5. 提交
+    const committed = await commitPatch(patchId);
+    if (!committed) {
+      console.log('[loop] commit 失败 (apply 或 lefthook 拦), 算 fail');
+      consecutiveFails++;
+      // commit 失败时只清 staged + 回滚 working tree (tracked), 不删 untracked 新文件
+      try {
+        await pExec('git', ['reset'], { cwd: REPO });
+        await pExec('git', ['checkout', '--', '.'], { cwd: REPO });
+      } catch { /* 忽略 */ }
+      continue;
+    }
+
+    // 6. 跑 vitest 看 commit 后是否真过
+    const after = await runVitest(targetFile);
+    console.log(`[vitest after] passed=${after.passed} failed=${after.failed}`);
+    if (after.failed < result.failed) {
+      console.log(`[loop] 进步: ${result.failed} → ${after.failed} fail`);
+      consecutiveFails = 0;
+    } else {
+      console.log(`[loop] 没进步, revert 这次 commit`);
+      await pExec('git', ['reset', '--hard', 'HEAD~1'], { cwd: REPO });
+      consecutiveFails++;
+    }
+
+    if (consecutiveFails >= 3) {
+      await notify(`连续 3 次无进步, 自动回滚 ${baseline}`);
+      await rollback(baseline);
+      return;
+    }
+  }
+
+  console.log(`[loop] 达到 max-iter=${maxIter}, 退出`);
+}
+
+main().catch((e) => { console.error('[loop] fatal:', e); process.exit(1); });

package/scripts/auto-evolve-oneshot.sh

@@ -0,0 +1,155 @@
+#!/bin/bash
+# auto-evolve-oneshot.sh — 阶段 D 单次修复 (shell 版)
+#
+# 流程 (全本地, 用 MINIMAX_API_KEY 调 LLM):
+#   1. 跑 vitest, 抓 fail
+#   2. 让 LLM 修 
+#   3. 解析 LLM 输出的 diff, 写到 staging
+#   4. 跑 reviewer (护栏 4)
+#   5. PASS → git apply + commit (护栏 1 拦)
+
+set -uo pipefail  # 不加 -e: vitest 失败要继续
+
+REPO="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$REPO"
+
+echo "[oneshot] REPO=$REPO"
+
+# 1. 跑 vitest
+npx vitest run --reporter=json --no-color 2>/dev/null > /tmp/vt-out.json
+TOTAL_FAIL=$(python3 -c "
+import json
+d = json.load(open('/tmp/vt-out.json'))
+print(sum(1 for f in d.get('testResults',[]) for a in f.get('assertionResults',[]) if a.get('status')=='failed'))
+")
+echo "[oneshot] vitest failed=$TOTAL_FAIL"
+
+if [ "$TOTAL_FAIL" = "0" ]; then
+  echo "[oneshot] ✅ 全部通过, 不需要修"
+  exit 0
+fi
+
+# 2. 抽 fail 信息
+FAIL_SUMMARY=$(python3 -c "
+import json
+d = json.load(open('/tmp/vt-out.json'))
+for f in d.get('testResults', []):
+    for a in f.get('assertionResults', []):
+        if a.get('status') == 'failed':
+            print('FILE:', f.get('name'))
+            print('TEST:', a.get('fullName') or a.get('title'))
+            print('ERROR:')
+            for m in a.get('failureMessages', [])[:2]:
+                print(m[:600])
+            print('---')
+" | head -60)
+echo "[oneshot] 2. 调 LLM 修..."
+
+# 3. 用 LLM 生成修复
+# prompt: 要 LLM 输出 ```diff ... ``` 块
+# 3. 写 prompt 到文件 (避免 shell 反引号冲突)
+PROMPT_FILE="/tmp/oneshot-prompt.txt"
+cat > "$PROMPT_FILE" <<PROMPT_EOF
+你是一个谨慎的代码修复助手. 你的工作是修复失败的测试.
+
+约束: 改动最小, 不改测试, 不引入 any/unknown/@ts-ignore.
+
+输出格式: 严格只输出一个 \`\`\`diff ... \`\`\` 块, 第一个字符 \`\`\`diff, 最后一个字符 \`\`\`. 中间是 unified diff (--- a/path +++ b/path 风格). 不要在 diff 块外输出任何文字.
+
+FAIL 信息:
+$(cat /tmp/vt-out.json | python3 -c "
+import json
+d = json.load(open('/tmp/vt-out.json'))
+for f in d.get('testResults', []):
+    for a in f.get('assertionResults', []):
+        if a.get('status') == 'failed':
+            print('FILE:', f.get('name'))
+            print('TEST:', a.get('fullName') or a.get('title'))
+            print('ERROR:')
+            for m in a.get('failureMessages', [])[:2]:
+                print(m[:600])
+            print('---')
+" | head -60)
+
+请**只**输出 diff 块:
+PROMPT_EOF
+
+LLM_OUTPUT=$(npx tsx -r dotenv/config -e "
+import { initMinimax } from './src/llm/pi-ai.js';
+import * as fs from 'fs';
+const prompt = fs.readFileSync('$PROMPT_FILE', 'utf-8');
+const client = initMinimax();
+const text = await client.generateText({ messages: [{ role: 'user', content: prompt }], maxTokens: 4096, temperature: 0.2 });
+process.stdout.write(text || '');
+" 2>/tmp/llm-err.log)
+
+if [ -z "$LLM_OUTPUT" ]; then
+  echo "[oneshot] ❌ LLM 没返回"
+  exit 2
+fi
+
+# 4. 解析 diff
+DIFF=$(echo "$LLM_OUTPUT" | python3 -c "
+import sys, re
+text = sys.stdin.read()
+m = re.search(r'\`\`\`diff\s*([\s\S]*?)\`\`\`', text)
+if m:
+    diff = m.group(1).strip()
+    if not diff.endswith('\n'):
+        diff += '\n'
+    print(diff)
+else:
+    sys.exit(1)
+" 2>/dev/null) || {
+  echo "[oneshot] ❌ LLM 输出没 diff 块"
+  echo "--- LLM 原始输出 (前 800) ---"
+  echo "$LLM_OUTPUT" | head -c 800
+  echo ""
+  exit 3
+}
+
+echo "[oneshot] 拿到 diff: $(echo "$DIFF" | wc -l) lines"
+
+# 5. 写 staging
+ID="oneshot-$(date +%s)"
+mkdir -p "staging/auto-evolve/$ID"
+echo "$DIFF" > "staging/auto-evolve/$ID/$ID.patch"
+echo "$ID" > "staging/auto-evolve/$ID/.patch-id"
+echo "[oneshot] patch 写到 staging/auto-evolve/$ID/"
+
+# 6. 跑 reviewer
+echo "[oneshot] 3. 跑 reviewer..."
+npx tsx -r dotenv/config scripts/diff-reviewer.ts "$ID" > /tmp/reviewer.log 2>&1 || true
+VERDICT=$(python3 -c "
+import json
+try:
+    print(json.load(open('staging/auto-evolve/$ID/.review-verdict')).get('verdict','UNKNOWN'))
+except:
+    print('UNKNOWN')
+")
+echo "[oneshot] reviewer verdict: $VERDICT"
+
+# 7. apply + commit
+if [ "$VERDICT" = "PASS" ]; then
+  echo "[oneshot] 4. apply + commit"
+  if git apply --recount --whitespace=fix "staging/auto-evolve/$ID/$ID.patch" 2>/tmp/apply.err; then
+    git add -A
+    git commit -m "auto-evolve: $ID (LLM 修复)"
+    echo "[oneshot] ✅ 提交成功"
+    # 验证
+    npx vitest run --reporter=json --no-color 2>/dev/null > /tmp/vt-after.json
+    AFTER_FAIL=$(python3 -c "
+import json
+print(sum(1 for f in json.load(open('/tmp/vt-after.json')).get('testResults',[]) for a in f.get('assertionResults',[]) if a.get('status')=='failed'))
+")
+    echo "[oneshot] 修复后 fail: $AFTER_FAIL (之前 $TOTAL_FAIL)"
+  else
+    echo "[oneshot] ❌ git apply 失败:"
+    cat /tmp/apply.err
+    exit 4
+  fi
+else
+  echo "[oneshot] ❌ reviewer verdict=$VERDICT, 不 apply"
+  cat /tmp/reviewer.log | head -10
+  exit 5
+fi

package/scripts/auto-evolve-snapshot.sh

@@ -0,0 +1,136 @@
+#!/bin/bash
+# auto-evolve-snapshot.sh — 阶段 C 护栏 2 + 3
+#
+# 用法 (LLM 改源码前必调):
+#   bash scripts/auto-evolve-snapshot.sh          # 打 baseline tag + 记当前 HEAD
+#   bash scripts/auto-evolve-snapshot.sh apply <patch-id>   # 人类批准后合并 staging → main
+#   bash scripts/auto-evolve-snapshot.sh list      # 列所有 baseline
+#   bash scripts/auto-evolve-snapshot.sh rollback <tag>     # 回滚到指定 baseline
+#
+# 流程:
+#   1. LLM 改之前调 snapshot: 当前 HEAD 打 auto-evolve-baseline-<ts> tag
+#   2. LLM 改 staging/auto-evolve/<patch-id>/  (不进 src/)
+#   3. 护栏 1 (lefthook) 跑 vitest + tsc, 坏就 abort
+#   4. 护栏 4 (reviewer hook) 审 diff, 通过才准 apply
+#   5. 人类调 apply: git apply staging/auto-evolve/<patch-id>/*.patch → src/
+#   6. 出问题: 调 rollback → git reset --hard <tag>
+
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$REPO_ROOT"
+
+STAGING_DIR="staging/auto-evolve"
+TAG_PREFIX="auto-evolve-baseline-"
+
+cmd="${1:-snapshot}"
+patch_id="${2:-}"
+
+case "$cmd" in
+  snapshot)
+    # 护栏 2: 打 baseline tag
+    if ! git diff --quiet HEAD 2>/dev/null; then
+      echo "❌ 有未提交改动. 先 git commit 或 git stash"
+      exit 1
+    fi
+    ts="$(date -u +%Y%m%dT%H%M%SZ)"
+    tag="${TAG_PREFIX}${ts}"
+    git tag -a "$tag" -m "auto-evolve baseline @ ${ts}" HEAD
+    echo "✅ baseline tag: $tag"
+    echo "$tag" > .last-auto-evolve-baseline
+    echo "   回滚命令: bash $0 rollback $tag"
+    ;;
+
+  prepare)
+    # LLM 改之前调: 创建 staging 目录
+    if [ -z "$patch_id" ]; then
+      echo "用法: $0 prepare <patch-id>"
+      exit 1
+    fi
+    mkdir -p "$STAGING_DIR/$patch_id"
+    echo "$patch_id" > "$STAGING_DIR/$patch_id/.patch-id"
+    echo "✅ staging 创建: $STAGING_DIR/$patch_id/"
+    echo "   LLM 改完后把 patch 放这里: $STAGING_DIR/$patch_id/*.patch"
+    ;;
+
+  apply)
+    # 人类批准: 把 staging 的 patch 合并到 src/
+    if [ -z "$patch_id" ]; then
+      echo "用法: $0 apply <patch-id>"
+      exit 1
+    fi
+    patch_dir="$STAGING_DIR/$patch_id"
+    if [ ! -d "$patch_dir" ]; then
+      echo "❌ staging 不存在: $patch_dir"
+      exit 1
+    fi
+    # 必须先 snapshot
+    if [ ! -f .last-auto-evolve-baseline ]; then
+      echo "❌ 没有 baseline. 先跑: $0 snapshot"
+      exit 1
+    fi
+    baseline="$(cat .last-auto-evolve-baseline)"
+
+    # 应用所有 patch
+    applied=0
+    for p in "$patch_dir"/*.patch; do
+      [ -e "$p" ] || continue
+      echo "  applying: $p"
+      if ! git apply --check "$p"; then
+        echo "❌ patch 不可用: $p (可能已应用过)"
+        exit 1
+      fi
+      git apply "$p"
+      applied=$((applied + 1))
+    done
+
+    if [ $applied -eq 0 ]; then
+      echo "❌ staging 里没有 .patch 文件"
+      exit 1
+    fi
+
+    echo "✅ 应用 $applied 个 patch"
+    echo "   建议现在跑: npm test  +  git commit -m 'auto-evolve: $patch_id'"
+    echo "   出问题回滚: $0 rollback $baseline"
+    ;;
+
+  rollback)
+    # 回滚到指定 baseline
+    if [ -z "$patch_id" ]; then
+      echo "用法: $0 rollback <tag>"
+      echo "可用的 baseline:"
+      git tag -l "${TAG_PREFIX}*" | sort -r | head -10
+      exit 1
+    fi
+    if ! git tag -l | grep -qx "$patch_id"; then
+      echo "❌ tag 不存在: $patch_id"
+      exit 1
+    fi
+    echo "⚠️  将回滚到 $patch_id (hard reset, 丢弃之后所有改动)"
+    read -p "确认? [y/N] " -n 1 -r
+    echo
+    if [[ $REPLY =~ ^[Yy]$ ]]; then
+      git reset --hard "$patch_id"
+      echo "✅ 回滚到 $patch_id"
+    else
+      echo "取消"
+    fi
+    ;;
+
+  list)
+    echo "auto-evolve baselines (最近 10):"
+    git tag -l "${TAG_PREFIX}*" | sort -r | head -10 | while read tag; do
+      msg="$(git tag -l --format='%(contents)' "$tag" | head -1)"
+      echo "  $tag — $msg"
+    done
+    if [ -f .last-auto-evolve-baseline ]; then
+      echo ""
+      echo "当前 baseline: $(cat .last-auto-evolve-baseline)"
+    fi
+    ;;
+
+  *)
+    echo "用法: $0 {snapshot|prepare <id>|apply <id>|rollback <tag>|list}"
+    exit 1
+    ;;
+esac