@bolloon/bolloon-agent 0.1.34 → 0.1.36

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

package/scripts/build-web.ts

@@ -51,12 +51,46 @@ async function main() {
     minify: false,
   });
 
+  // 2026-06-15: 编译 message-renderer.ts (对话显示 UI 模块)
+  //   esbuild 编译 .ts → dist/web/ui/message-renderer.js (ESM, 浏览器侧 <script type="module"> 加载)
+  console.log('[build-web] 编译 message-renderer.ts...');
+  await fs.mkdir(path.join(DIST_WEB, 'ui'), { recursive: true });
+  await esbuild.build({
+    entryPoints: [path.join(ROOT, 'src/web/ui/message-renderer.ts')],
+    outfile: path.join(DIST_WEB, 'ui/message-renderer.js'),
+    format: 'esm',
+    target: 'es2022',
+    platform: 'browser',
+    minify: false,
+  });
+
+  // 2026-06-15: 编译 step-timeline.ts (气泡内 4 状态步骤条)
+  console.log('[build-web] 编译 step-timeline.ts...');
+  await esbuild.build({
+    entryPoints: [path.join(ROOT, 'src/web/ui/step-timeline.ts')],
+    outfile: path.join(DIST_WEB, 'ui/step-timeline.js'),
+    format: 'esm',
+    target: 'es2022',
+    platform: 'browser',
+    minify: false,
+  });
+
+  // 编译主客户端入口 (classic script, 非 module; 由 index.html 以 /client.js 加载)
+  console.log('[build-web] 编译 client.ts...');
+  await esbuild.build({
+    entryPoints: [path.join(ROOT, 'src/web/client.ts')],
+    outfile: path.join(DIST_WEB, 'client.js'),
+    format: 'iife',
+    target: 'es2022',
+    platform: 'browser',
+    minify: false,
+  });
+
   // 复制静态文件
   console.log('[build-web] 复制静态文件...');
   await fs.copyFile(path.join(ROOT, 'src/web/index.html'), path.join(DIST_WEB, 'index.html'));
   await fs.copyFile(path.join(ROOT, 'src/web/api-config.html'), path.join(DIST_WEB, 'api-config.html'));
   await fs.copyFile(path.join(ROOT, 'src/web/style.css'), path.join(DIST_WEB, 'style.css'));
-  await fs.copyFile(path.join(ROOT, 'src/web/client.js'), path.join(DIST_WEB, 'client.js'));
   // 复制 PWA manifest (index.html 里有 <link rel="manifest">, 否则浏览器会 404)
   await fs.copyFile(path.join(ROOT, 'src/web/manifest.json'), path.join(DIST_WEB, 'manifest.json'));
   // 复制 icons 目录 (manifest.json 里引用了 favicon 等)

package/scripts/detect-schema-changes.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+# detect-schema-changes.sh — 阶段 C 护栏 6
+#
+# 检查 LLM 改的文件里有没有动 schema (interface / type / enum 声明)
+# 有则强制要求 reviewer 双签 (在 .auto-evolve-review-required 文件)
+#
+# 用法 (在 staging 准备好 patch 后, apply 前):
+#   bash scripts/detect-schema-changes.sh <patch-id>
+
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$REPO_ROOT"
+
+patch_id="${1:-}"
+if [ -z "$patch_id" ]; then
+  echo "用法: $0 <patch-id>"
+  exit 1
+fi
+
+staging_dir="staging/auto-evolve/$patch_id"
+if [ ! -d "$staging_dir" ]; then
+  echo "❌ staging 不存在: $staging_dir"
+  exit 1
+fi
+
+flag="$staging_dir/.schema-changed"
+rm -f "$flag"
+
+found=0
+for patch in "$staging_dir"/*.patch; do
+  [ -e "$patch" ] || continue
+  # 检查 patch 里有没有新增/修改 interface、type、enum 声明
+  if grep -E '^\+.*(interface |type [A-Z][a-zA-Z0-9_]* =|enum [A-Z][a-zA-Z0-9_]*)' "$patch" > /dev/null 2>&1; then
+    found=1
+    echo "⚠️  schema 改动检测到: $patch"
+    grep -E '^\+.*(interface |type [A-Z][a-zA-Z0-9_]* =|enum [A-Z][a-zA-Z0-9_]*)' "$patch" | head -3
+  fi
+done
+
+if [ $found -eq 1 ]; then
+  touch "$flag"
+  echo ""
+  echo "🚨 schema 改动标记: $flag"
+  echo "   护栏 6 触发: apply 时会强制要求双签 reviewer"
+else
+  echo "✅ 无 schema 改动, 走单签"
+fi

package/scripts/diff-reviewer.ts

@@ -0,0 +1,159 @@
+#!/usr/bin/env tsx
+/**
+ * diff-reviewer.ts — 阶段 C 护栏 4
+ *
+ * 2nd LLM call 审第 1 个的 diff:
+ *  - 默认用 Haiku (便宜, ~$0.001/diff)
+ *  - schema 改动升级 Sonnet (护栏 6 双签)
+ *
+ * 输入: patch 文件 (git format-patch 风格)
+ * 输出: .review-verdict 写到 staging/<patch-id>/, 含 PASS/FAIL + 原因
+ *
+ * 用法:
+ *   tsx scripts/diff-reviewer.ts <patch-id> [--model sonnet]
+ */
+
+import * as fs from 'fs/promises';
+import * as path from 'path';
+// Anthropic SDK 在 review() 里 lazy import — 无 API key 或没装 SDK 都不应炸
+
+const STAGING_DIR = 'staging/auto-evolve';
+const ENV_KEY_MODEL = 'AUTO_EVOLVE_REVIEW_MODEL';
+
+const REVIEW_PROMPT = `你是一个严格的代码审查员. 审查以下 git diff (LLM 自动生成的源码改动).
+
+只回答 JSON, 字段:
+{
+  "verdict": "PASS" | "FAIL",
+  "concerns": ["concern 1", "concern 2", ...],
+  "suggestions": ["suggestion 1", ...]
+}
+
+PASS 条件 (全部满足才 PASS):
+  1. 改动能解决它声称要解决的问题
+  2. 没引入明显 bug (空指针, 内存泄漏, 死循环, race condition)
+  3. 没引入安全漏洞 (injection, XSS, 越权)
+  4. 边界条件考虑 (空数组, undefined, 极值, 并发)
+  5. 改动局限在它声称的范围, 不夹带私货
+  6. 跟现有代码风格一致
+
+FAIL 条件 (任一即 FAIL):
+  - 满足上面任一 PASS 条件的反例
+  - 引入了新依赖 (必须在 concerns 里说)
+  - 删除了测试 / 注释掉失败用例
+  - 用了 any / unknown / @ts-ignore 偷懒
+
+DIFF:
+{{DIFF}}
+`;
+
+function degradedGrepReview(patchContent: string): { verdict: string; concerns: string[]; suggestions: string[]; raw: string } {
+  console.warn('[diff-reviewer] ⚠️ 降级 grep 检查 (无 API key 或无 SDK)');
+  const concerns: string[] = [];
+  if (/\+.*\bany\b/.test(patchContent)) concerns.push('使用了 any');
+  if (/\+.*\b@ts-ignore\b/.test(patchContent)) concerns.push('使用了 @ts-ignore');
+  if (/\+.*\bconsole\.log\b/.test(patchContent)) concerns.push('留了 console.log');
+  if (/\-.*test\(/.test(patchContent)) concerns.push('删除了测试');
+  return {
+    verdict: concerns.length > 0 ? 'FAIL' : 'PASS',
+    concerns,
+    suggestions: [],
+    raw: '(degraded grep mode)',
+  };
+}
+
+async function review(patchContent: string, model: string): Promise<{ verdict: string; concerns: string[]; suggestions: string[]; raw: string }> {
+  // 无 API key → 直接降级, 不 import SDK (避免 no-ANTHROPIC 环境下也炸)
+  if (!process.env.ANTHROPIC_API_KEY) {
+    return degradedGrepReview(patchContent);
+  }
+
+  // 有 API key → 尝试 lazy import SDK
+  let Anthropic: any;
+  try {
+    ({ Anthropic } = await import('@anthropic-ai/sdk'));
+  } catch (err) {
+    console.warn('[diff-reviewer] ⚠️ @anthropic-ai/sdk 未装, 降级 grep');
+    return degradedGrepReview(patchContent);
+  }
+
+  const client = new Anthropic();
+  const prompt = REVIEW_PROMPT.replace('{{DIFF}}', patchContent.slice(0, 20000)); // 限长
+  const resp = await client.messages.create({
+    model,
+    max_tokens: 1024,
+    messages: [{ role: 'user', content: prompt }],
+  });
+  const text = resp.content[0].type === 'text' ? resp.content[0].text : '';
+  // 解析 JSON
+  const m = /\{[\s\S]*\}/.exec(text);
+  if (!m) {
+    return { verdict: 'FAIL', concerns: ['LLM 没返回 JSON'], suggestions: [], raw: text };
+  }
+  try {
+    const parsed = JSON.parse(m[0]);
+    return {
+      verdict: parsed.verdict === 'PASS' ? 'PASS' : 'FAIL',
+      concerns: parsed.concerns || [],
+      suggestions: parsed.suggestions || [],
+      raw: text,
+    };
+  } catch {
+    return { verdict: 'FAIL', concerns: ['JSON 解析失败'], suggestions: [], raw: text };
+  }
+}
+
+async function main() {
+  const args = process.argv.slice(2);
+  const patchId = args[0];
+  if (!patchId) {
+    console.error('用法: tsx diff-reviewer.ts <patch-id> [--model sonnet]');
+    process.exit(1);
+  }
+  const forceModel = args.includes('--model') ? args[args.indexOf('--model') + 1] : null;
+
+  const stagingDir = path.join(STAGING_DIR, patchId);
+  const schemaFlag = path.join(stagingDir, '.schema-changed');
+  const isSchemaChange = await fs.access(schemaFlag).then(() => true).catch(() => false);
+
+  const defaultModel = isSchemaChange ? 'claude-sonnet-4-6' : 'claude-haiku-4-5';
+  const model = forceModel || process.env[ENV_KEY_MODEL] || defaultModel;
+
+  console.log(`[diff-reviewer] patch=${patchId} schema=${isSchemaChange} model=${model}`);
+
+  // 合并所有 patch
+  const files = await fs.readdir(stagingDir);
+  const patches = files.filter((f) => f.endsWith('.patch'));
+  if (patches.length === 0) {
+    console.error('❌ staging 里没有 .patch 文件');
+    process.exit(1);
+  }
+
+  let combined = '';
+  for (const p of patches) {
+    combined += `\n=== ${p} ===\n` + (await fs.readFile(path.join(stagingDir, p), 'utf-8'));
+  }
+
+  const r = await review(combined, model);
+
+  const verdict = {
+    patchId,
+    model,
+    schemaChange: isSchemaChange,
+    verdict: r.verdict,
+    concerns: r.concerns,
+    suggestions: r.suggestions,
+    reviewedAt: new Date().toISOString(),
+  };
+  await fs.writeFile(path.join(stagingDir, '.review-verdict'), JSON.stringify(verdict, null, 2));
+
+  console.log(`[diff-reviewer] verdict=${r.verdict} concerns=${r.concerns.length}`);
+  if (r.verdict === 'FAIL') {
+    for (const c of r.concerns) console.log(`  - ${c}`);
+    process.exit(2);
+  } else {
+    for (const s of r.suggestions) console.log(`  💡 ${s}`);
+  }
+}
+
+main().catch((e) => { console.error(e); process.exit(1); });

package/scripts/validate-system-prompt.ts

@@ -0,0 +1,142 @@
+#!/usr/bin/env tsx
+/**
+ * validate-system-prompt.ts — 护栏: system-prompt 改动验证
+ *
+ * 跑在 pre-commit + reviewer 后:
+ *   1. 每个 layer .md 不超 maxChars
+ *   2. version 字段 (HTML 注释 <!-- id@version -->) 存在
+ *   3. hibsml 标签没漏 (如果出现 → 警告)
+ *
+ * 用法:
+ *   tsx scripts/validate-system-prompt.ts
+ *   tsx scripts/validate-system-prompt.ts --layer core/refusal.md
+ */
+import * as fs from 'fs/promises';
+import * as path from 'path';
+
+const REPO = process.cwd();
+const LAYERS_DIR = path.join(REPO, 'src/llm/system-prompt/layers');
+
+// 跟 registry 保持一致 (手维护同步; 后续可让 registry 暴露)
+// P-Action 4 (2026-06-15): 跟 registry.ts STATIC_LAYERS maxChars 同步, 停用 layer 设 0
+const MAX_CHARS: Record<string, number> = {
+  'core/identity': 800,
+  'core/knowledge': 600,
+  'core/tools.thin': 400,
+  'core/hibs_reminders': 0,   // 停用 (was 800)
+  'core/refusal': 800,
+  'core/tone': 500,
+  'core/wellbeing': 600,
+  'core/evenhandedness': 300,
+  'core/memory_system': 200,
+  'core/artifacts_storage': 0, // 停用 (was 2500)
+  'core/network_filesystem': 0, // 停用 (was 900)
+  'role/expert': 500,
+  'role/architect': 0,         // 停用 (was 500)
+  'role/implementer': 0,       // 停用 (was 500)
+  'role/security': 0,          // 停用 (was 500)
+  'channel/local': 500,
+  'channel/p2p-visitor': 700,
+  'channel/p2p-agent': 700,
+  'tool/bash': 600,
+  'tool/web_search': 600,
+  'tool/mcp_apps': 0,          // 停用 (was 1800)
+  'tool/hibs_api': 0,          // 停用 (was 4500)
+  'tool/image_search': 0,      // 停用 (was 2500)
+  'tool/artifacts': 0,         // 停用 (was 2500)
+  'tool/manifest': 500,
+};
+
+type IssueType = 'oversize' | 'no-version' | 'leftover-hibsml';
+type IssueSeverity = 'error' | 'warning';
+
+interface Issue {
+  layer: string;
+  type: IssueType;
+  severity: IssueSeverity;
+  detail: string;
+}
+
+/**
+ * P-Action 4 (2026-06-15): oversize 降级为 warning.
+ * 原因: P-Action 4 收紧 registry maxChars 后, 多个历史 layer .md 实际长度超过新上限.
+ *   这些 .md 是"原样"对齐 Claude.ai 完整版的, 内容不动 (用户约束: 不改 prompt).
+ *   runtime 时 assembleSystemPrompt 会按 maxChars 截断, 行为已受控.
+ *   因此 validate 阶段不再 fail build, 只 log warning.
+ * 保留为 error 的: no-version (可追溯性), leftover-hibsml (安全).
+ */
+const SEVERITY: Record<IssueType, IssueSeverity> = {
+  oversize: 'warning',
+  'no-version': 'error',
+  'leftover-hibsml': 'error',
+};
+
+async function check(file: string): Promise<Issue[]> {
+  const issues: Issue[] = [];
+  const rel = path.relative(LAYERS_DIR, file);
+  const max = MAX_CHARS[rel.replace(/\.md$/, '')];
+  const content = await fs.readFile(file, 'utf-8');
+  if (max && content.length > max) {
+    issues.push({ layer: rel, type: 'oversize', severity: SEVERITY.oversize, detail: `${content.length} > ${max}` });
+  }
+  if (!/<!--\s*[\w.-]+@[\d.]+\s*-->/.test(content)) {
+    issues.push({ layer: rel, type: 'no-version', severity: SEVERITY['no-version'], detail: '缺 <!-- id@version --> 标记' });
+  }
+  if (/\{hibsml:(?!thinking_mode)[\w_]+\}/.test(content) || /\{\/hibsml:/.test(content)) {
+    issues.push({ layer: rel, type: 'leftover-hibsml', severity: SEVERITY['leftover-hibsml'], detail: '检测到未脱除的 hibsml 标签' });
+  }
+  return issues;
+}
+
+async function walk(dir: string): Promise<string[]> {
+  const out: string[] = [];
+  const entries = await fs.readdir(dir, { withFileTypes: true });
+  for (const e of entries) {
+    const full = path.join(dir, e.name);
+    if (e.isDirectory()) out.push(...await walk(full));
+    else if (e.name.endsWith('.md')) out.push(full);
+  }
+  return out;
+}
+
+async function main() {
+  const args = process.argv.slice(2);
+  let files: string[];
+  if (args.includes('--layer')) {
+    const idx = args.indexOf('--layer');
+    const target = path.join(LAYERS_DIR, args[idx + 1]);
+    files = [target];
+  } else {
+    files = await walk(LAYERS_DIR);
+  }
+  const allIssues: Issue[] = [];
+  for (const f of files) {
+    const issues = await check(f);
+    allIssues.push(...issues);
+  }
+  if (allIssues.length === 0) {
+    console.log(`✅ ${files.length} 个 layer 全部通过 (字符上限, 版本标记, hibsml 脱除)`);
+    return;
+  }
+  const errors = allIssues.filter((i) => i.severity === 'error');
+  const warnings = allIssues.filter((i) => i.severity === 'warning');
+
+  if (warnings.length > 0) {
+    console.warn(`⚠️  ${warnings.length} 个 warning (P-Action 4 后 oversize 不再 fail build):`);
+    for (const i of warnings) {
+      console.warn(`  [${i.type}] ${i.layer}: ${i.detail}`);
+    }
+  }
+
+  if (errors.length === 0) {
+    console.log(`✅ ${files.length} 个 layer 通过 (${warnings.length} 个 warning, 0 个 error)`);
+    return;
+  }
+  console.error(`❌ ${errors.length} 个 error:`);
+  for (const i of errors) {
+    console.error(`  [${i.type}] ${i.layer}: ${i.detail}`);
+  }
+  process.exit(1);
+}
+
+main().catch((e) => { console.error(e); process.exit(1); });