@bolloon/bolloon-agent 0.1.34 → 0.1.36

package/scripts/weekly-report.ts

@@ -0,0 +1,364 @@
+#!/usr/bin/env tsx
+/**
+ * weekly-report.ts — Bolloon 每周表现报告 (阶段 B)
+ *
+ * 输入:
+ *   ~/.bolloon/human-values/evolution.jsonl      (接受/拒绝/回滚事件)
+ *   ~/.bolloon/human-values/usage.jsonl          (judgment 使用记录)
+ *   ~/.bolloon/human-values/counterfactual-audit.jsonl (反事实审计)
+ *   ~/.bolloon/human-values/judgments.json       (当前 judgment 库)
+ *   ~/.bolloon/self-improve-audit.log            (改源码尝试的审计, 即使被拒)
+ *
+ * 输出:
+ *   ~/.bolloon/reports/2026-W24.md  (markdown 报告)
+ *
+ * 设计原则:
+ *   - 纯本地计算 + 纯函数式分析 (不调 LLM, 避免幻觉)
+ *   - 周范围默认 ISO 周(周一为起点), 可 --week 2026-W24 指定
+ *   - 不写 judgments.json, 不动 persona, 仅追加 reports/*.md
+ *   - 失败静默 + 退出码 != 0 让 cron 知道坏了
+ *
+ * 用法:
+ *   tsx scripts/weekly-report.ts                # 生成上周
+ *   tsx scripts/weekly-report.ts --week 2026-W24  # 指定周
+ *   tsx scripts/weekly-report.ts --week 2026-W24 --dry-run
+ */
+
+import * as fs from 'fs/promises';
+import * as os from 'os';
+import * as path from 'path';
+
+const HOME = () => os.homedir() || process.env.HOME || '/tmp';
+const ROOT = () => HOME() + '/.bolloon';
+const REPORTS_DIR = () => ROOT() + '/reports';
+
+const FILES = {
+  evolution: () => ROOT() + '/human-values/evolution.jsonl',
+  usage: () => ROOT() + '/human-values/usage.jsonl',
+  counterfactual: () => ROOT() + '/human-values/counterfactual-audit.jsonl',
+  judgments: () => ROOT() + '/human-values/judgments.json',
+  selfImproveAudit: () => ROOT() + '/self-improve-audit.log',
+};
+
+interface EvolutionEntry {
+  ts: string;
+  action: 'accept' | 'reject' | 'revert';
+  suggestion: { kind: string; judgmentId: string; action: string };
+  appliedPatch?: Record<string, unknown>;
+}
+interface UsageEntry {
+  ts: string;
+  channelId: string | null;
+  userInputPreview: string;
+  usedIds: string[];
+}
+interface CounterfactualEntry {
+  ts: string;
+  trigger: { userInput: string; aiReply: string; violatedPrinciples: unknown[] };
+  verdict: string;
+  recomendaciones?: string[];
+}
+
+async function readJsonl<T>(p: string): Promise<T[]> {
+  try {
+    const content = await fs.readFile(p, 'utf-8');
+    return content
+      .trim()
+      .split('\n')
+      .filter(Boolean)
+      .map((l) => {
+        try {
+          return JSON.parse(l) as T;
+        } catch {
+          return null;
+        }
+      })
+      .filter((e): e is T => Boolean(e));
+  } catch {
+    return [];
+  }
+}
+
+async function readJson<T>(p: string): Promise<T | null> {
+  try {
+    return JSON.parse(await fs.readFile(p, 'utf-8')) as T;
+  } catch {
+    return null;
+  }
+}
+
+function isoWeekString(d: Date): string {
+  // ISO week: 1 = 包含 1 月 4 日的那周
+  const target = new Date(d.valueOf());
+  const dayNr = (d.getUTCDay() + 6) % 7; // 周一=0
+  target.setUTCDate(target.getUTCDate() - dayNr + 3);
+  const firstThursday = new Date(Date.UTC(target.getUTCFullYear(), 0, 4));
+  const week =
+    1 +
+    Math.round(
+      ((target.valueOf() - firstThursday.valueOf()) / 86400000 - 3 + ((firstThursday.getUTCDay() + 6) % 7)) / 7
+    );
+  return `${target.getUTCFullYear()}-W${String(week).padStart(2, '0')}`;
+}
+
+function weekRange(iso: string): { start: Date; end: Date } {
+  const m = /^(\d{4})-W(\d{1,2})$/.exec(iso);
+  if (!m) throw new Error(`bad ISO week: ${iso}`);
+  const year = Number(m[1]);
+  const week = Number(m[2]);
+  // ISO 周 1 是包含 1 月 4 日的那周
+  const jan4 = new Date(Date.UTC(year, 0, 4));
+  const jan4Day = (jan4.getUTCDay() + 6) % 7; // 周一=0
+  const week1Mon = new Date(jan4);
+  week1Mon.setUTCDate(jan4.getUTCDate() - jan4Day);
+  const start = new Date(week1Mon);
+  start.setUTCDate(week1Mon.getUTCDate() + (week - 1) * 7);
+  const end = new Date(start);
+  end.setUTCDate(start.getUTCDate() + 7);
+  return { start, end };
+}
+
+function inWeek(ts: string, start: Date, end: Date): boolean {
+  const t = new Date(ts).getTime();
+  return t >= start.getTime() && t < end.getTime();
+}
+
+function pct(n: number, d: number): string {
+  if (d === 0) return 'n/a';
+  return `${((n / d) * 100).toFixed(1)}%`;
+}
+
+function topN<T>(arr: T[], key: (x: T) => string, n: number): Array<[string, number]> {
+  const m = new Map<string, number>();
+  for (const x of arr) {
+    const k = key(x);
+    m.set(k, (m.get(k) || 0) + 1);
+  }
+  return Array.from(m.entries())
+    .sort((a, b) => b[1] - a[1])
+    .slice(0, n);
+}
+
+interface Report {
+  week: string;
+  range: { start: string; end: string };
+  generatedAt: string;
+  summary: {
+    totalUsage: number;
+    uniqueChannels: number;
+    uniqueJudgments: number;
+    evolutionEvents: number;
+    acceptRate: number;
+    rejected: number;
+    reverted: number;
+    counterfactualScans: number;
+  };
+  topJudgments: Array<[string, number]>;
+  topKinds: Array<[string, number]>;
+  policyAudit: {
+    selfImproveAttempts: number;
+    blockedByPolicy: number;
+    note: string;
+  };
+  openQuestions: string[];
+}
+
+async function buildReport(weekIso: string): Promise<Report> {
+  const { start, end } = weekRange(weekIso);
+
+  const [evolution, usage, counterfactual, judgments, audit] = await Promise.all([
+    readJsonl<EvolutionEntry>(FILES.evolution()),
+    readJsonl<UsageEntry>(FILES.usage()),
+    readJsonl<CounterfactualEntry>(FILES.counterfactual()),
+    readJson<unknown[]>(FILES.judgments()),
+    (async () => {
+      try {
+        const txt = await fs.readFile(FILES.selfImproveAudit(), 'utf-8');
+        return txt.split('\n').filter(Boolean);
+      } catch {
+        return [];
+      }
+    })(),
+  ]);
+
+  const evoInWeek = evolution.filter((e) => inWeek(e.ts, start, end));
+  const useInWeek = usage.filter((u) => inWeek(u.ts, start, end));
+  const cfInWeek = counterfactual.filter((c) => inWeek(c.ts, start, end));
+  const auditInWeek = audit.filter((line) => {
+    // audit 是日志行, 格式不一定, 简单按日期前缀过滤
+    const d = line.match(/^(\d{4}-\d{2}-\d{2})/);
+    if (!d) return false;
+    const ts = new Date(d[1] + 'T00:00:00Z').getTime();
+    return ts >= start.getTime() && ts < end.getTime();
+  });
+
+  const totalUsage = useInWeek.length;
+  const channels = new Set(useInWeek.map((u) => u.channelId || 'null'));
+  const ids = new Set<string>();
+  for (const u of useInWeek) for (const id of u.usedIds) ids.add(id);
+
+  const accepts = evoInWeek.filter((e) => e.action === 'accept').length;
+  const rejects = evoInWeek.filter((e) => e.action === 'reject').length;
+  const reverts = evoInWeek.filter((e) => e.action === 'revert').length;
+  const acceptRate = accepts + rejects === 0 ? 0 : accepts / (accepts + rejects);
+
+  // 自改审计 (路径白/黑名单拦截)
+  const selfImproveAttempts = auditInWeek.filter((l) => /attempt|尝试/i.test(l)).length;
+  const blockedByPolicy = auditInWeek.filter((l) => /block|deny|拒绝|denylist/i.test(l)).length;
+
+  return {
+    week: weekIso,
+    range: { start: start.toISOString().slice(0, 10), end: end.toISOString().slice(0, 10) },
+    generatedAt: new Date().toISOString(),
+    summary: {
+      totalUsage,
+      uniqueChannels: channels.size,
+      uniqueJudgments: ids.size,
+      evolutionEvents: evoInWeek.length,
+      acceptRate,
+      rejected: rejects,
+      reverted: reverts,
+      counterfactualScans: cfInWeek.length,
+    },
+    topJudgments: topN(useInWeek.flatMap((u) => u.usedIds.map((id) => ({ id }))), (x) => x.id, 10),
+    topKinds: topN(evoInWeek, (e) => e.suggestion.kind, 5),
+    policyAudit: {
+      selfImproveAttempts,
+      blockedByPolicy,
+      note: selfImproveAttempts === 0 ? '本周无源码自改尝试(护栏未触发)' : '见 self-improve-audit.log',
+    },
+    openQuestions: openQuestions(evoInWeek, cfInWeek, useInWeek),
+  };
+}
+
+function openQuestions(evo: EvolutionEntry[], cf: CounterfactualEntry[], use: UsageEntry[]): string[] {
+  const out: string[] = [];
+  if (evo.length === 0 && use.length > 0) {
+    out.push('本周有使用但无自适应建议 → 可能 judgment 库太稳定, 或扫描器未触发');
+  }
+  if (use.length === 0) {
+    out.push('本周无 judgment 使用记录 → 检查 usage.jsonl 是否在写, 或渠道是否活跃');
+  }
+  if (cf.length > 0) {
+    const conflictCount = cf.filter((c) => /冲突|conflict|不合理/i.test(c.verdict)).length;
+    if (conflictCount > 0) {
+      out.push(`反事实审计发现 ${conflictCount} 条潜在冲突 → 看 counterfactual-audit.jsonl`);
+    }
+  }
+  const reverts = evo.filter((e) => e.action === 'revert').length;
+  if (reverts >= 2) {
+    out.push(`本周回滚 ${reverts} 次 → 类 B 建议可能过激, 考虑收紧阈值`);
+  }
+  return out;
+}
+
+function toMarkdown(r: Report, totalJudgments: number): string {
+  const lines: string[] = [];
+  lines.push(`# 📊 Bolloon 周报 — ${r.week}`);
+  lines.push('');
+  lines.push(`> 范围: ${r.range.start} → ${r.range.end} · 生成于 ${r.generatedAt}`);
+  lines.push('');
+  lines.push('## 核心数字');
+  lines.push('');
+  lines.push('| 指标 | 本周 |');
+  lines.push('|------|------|');
+  lines.push(`| judgment 调用次数 | ${r.summary.totalUsage} |`);
+  lines.push(`| 触达渠道数 | ${r.summary.uniqueChannels} |`);
+  lines.push(`| 用到的不同 judgment 数 | ${r.summary.uniqueJudgments} |`);
+  lines.push(`| 自适应事件数 | ${r.summary.evolutionEvents} |`);
+  lines.push(`| 接受率 | ${pct(r.summary.acceptRate * 100, 100)} (${Math.round(r.summary.acceptRate * 100)}%) |`);
+  lines.push(`| 拒绝数 | ${r.summary.rejected} |`);
+  lines.push(`| 回滚数 | ${r.summary.reverted} |`);
+  lines.push(`| 反事实审计次数 | ${r.summary.counterfactualScans} |`);
+  lines.push(`| 当前 judgment 库总条数 | ${totalJudgments} |`);
+  lines.push('');
+  lines.push('## 最常被引用的 judgment');
+  lines.push('');
+  if (r.topJudgments.length === 0) {
+    lines.push('_(本周无引用)_');
+  } else {
+    for (const [id, n] of r.topJudgments) {
+      lines.push(`- \`${id}\` × ${n}`);
+    }
+  }
+  lines.push('');
+  lines.push('## 自适应建议类型分布');
+  lines.push('');
+  if (r.topKinds.length === 0) {
+    lines.push('_(本周无自适应事件)_');
+  } else {
+    for (const [kind, n] of r.topKinds) {
+      lines.push(`- **${kind}** × ${n}`);
+    }
+  }
+  lines.push('');
+  lines.push('## 护栏审计');
+  lines.push('');
+  lines.push(`- 源码自改尝试: **${r.policyAudit.selfImproveAttempts}**`);
+  lines.push(`- 被策略拦截: **${r.policyAudit.blockedByPolicy}**`);
+  lines.push(`- ${r.policyAudit.note}`);
+  lines.push('');
+  lines.push('## 关注事项');
+  lines.push('');
+  if (r.openQuestions.length === 0) {
+    lines.push('_(本周一切正常, 无特别关注)_');
+  } else {
+    for (const q of r.openQuestions) {
+      lines.push(`- ${q}`);
+    }
+  }
+  lines.push('');
+  lines.push('---');
+  lines.push('');
+  lines.push('> 本报告由 `scripts/weekly-report.ts` 纯本地生成, 不调 LLM. 数据源见 ~/.bolloon/human-values/');
+  lines.push('');
+  return lines.join('\n');
+}
+
+function parseArgs(argv: string[]): { week?: string; dryRun: boolean } {
+  const out: { week?: string; dryRun: boolean } = { dryRun: false };
+  for (let i = 0; i < argv.length; i++) {
+    if (argv[i] === '--week' && argv[i + 1]) {
+      out.week = argv[++i];
+    } else if (argv[i] === '--dry-run') {
+      out.dryRun = true;
+    }
+  }
+  return out;
+}
+
+async function main() {
+  const args = parseArgs(process.argv.slice(2));
+  // 默认: 上周(更合理 — 周末才回顾)
+  const now = new Date();
+  const lastWeek = new Date(now);
+  lastWeek.setUTCDate(now.getUTCDate() - 7);
+  const week = args.week || isoWeekString(lastWeek);
+  const { start, end } = weekRange(week);
+
+  console.log(`[weekly-report] 生成 ${week} (${start.toISOString().slice(0, 10)} → ${end.toISOString().slice(0, 10)})`);
+
+  const r = await buildReport(week);
+  const judgments = (await readJson<unknown[]>(FILES.judgments())) || [];
+  const md = toMarkdown(r, judgments.length);
+
+  const outDir = REPORTS_DIR();
+  const outPath = path.join(outDir, `${week}.md`);
+
+  if (args.dryRun) {
+    console.log(`[weekly-report] DRY-RUN, 不会写盘. 输出预览:`);
+    console.log('---');
+    console.log(md);
+    console.log('---');
+    return;
+  }
+
+  await fs.mkdir(outDir, { recursive: true });
+  await fs.writeFile(outPath, md, 'utf-8');
+  console.log(`[weekly-report] ✅ 写入 ${outPath} (${md.length} bytes)`);
+}
+
+main().catch((err) => {
+  console.error('[weekly-report] ❌ 失败:', err);
+  process.exit(1);
+});

package/src/agents/p2p-chat-tools.ts

@@ -1,7 +1,7 @@
 /**
  * p2p-chat-tools — 异步 chat 通道 + 持久化 inbox + 判断力外包 (draft)
  *
- * 解决痛点: 两个节点的智能体代替各自人类主人做异步判断
+ * 解决痛点: 两个节点的智能体代替各自人类用户做异步判断
  *
  * 流程:
  *   1. A 节点 (人类在线) 通过 sendChat(peerDID, text) 发消息 → iroh 'agent_chat'
@@ -9,7 +9,7 @@
  *        - onMessage('agent_chat') → 落 ~/.bolloon/inbox/<peerDID>.jsonl
  *        - 状态 = 'received' (未处理)
  *   3. B 节点 wake-up (processPendingInbox) → 扫描 status='received' 的消息
- *        - 调 LLM 生成 draft (注入主人历史判断 + ValueProfile)
+ *        - 调 LLM 生成 draft (注入用户历史判断 + ValueProfile)
  *        - 写 status='drafted', draft 落盘
  *        - 通过 'agent_chat' 消息类型回送 draft (前缀 [DRAFT] 表明是代回)
  *   4. A 节点收到 draft → 写到 A 的 outbox (inbox 中 from=B 的那条)
@@ -258,7 +258,7 @@ async function buildValueHint(text: string): Promise<string> {
       .slice(-10)
       .flatMap((j) => j.reasons || [])
       .slice(0, 20);
-    return `\n[主人历史判断 (style 参考, 不可外泄)]\n关注维度: ${profileHint}\n关键理由示例: ${reasons.join(' | ').slice(0, 400) || '(无)'}\n`;
+    return `\n[用户历史判断 (style 参考, 不可外泄)]\n关注维度: ${profileHint}\n关键理由示例: ${reasons.join(' | ').slice(0, 400) || '(无)'}\n`;
   } catch (e) {
     return '';
   }
@@ -275,7 +275,7 @@ export async function generateDraft(
   if (entry.status !== 'received') return entry;
 
   const valueHint = await buildValueHint(entry.text);
-  const promptForDraft = `你是主人的代理. 对方发来这条消息: "${entry.text.slice(0, 1500)}"\n${valueHint}\n请基于主人的历史判断, 用 1-2 句话代主人拟一个回复草案. 草案要保留主人的语气和立场, 开头标注 [DRAFT]. 直接给草案文本, 不要解释.`;
+  const promptForDraft = `你是用户的代理. 对方发来这条消息: "${entry.text.slice(0, 1500)}"\n${valueHint}\n请基于用户的历史判断, 用 1-2 句话代用户拟一个回复草案. 草案要保留用户的语气和立场, 开头标注 [DRAFT]. 直接给草案文本, 不要解释.`;
 
   let draftText = '';
   let confidence = 0.5;
@@ -290,7 +290,7 @@ export async function generateDraft(
       body: JSON.stringify({
         model: openaiModel,
         messages: [
-          { role: 'system', content: '你是主人的代理. 你的输出会被主人审阅后才发出. 请谨慎.' },
+          { role: 'system', content: '你是用户的代理. 你的输出会被用户审阅后才发出. 请谨慎.' },
           { role: 'user', content: promptForDraft },
         ],
         temperature: 0.4,
@@ -310,7 +310,7 @@ export async function generateDraft(
   }
 
   if (!draftText) {
-    draftText = `[DRAFT] 已收到. (本地 draft 引擎未配置 LLM, 主人上线后请手写回复)`;
+    draftText = `[DRAFT] 已收到. (本地 draft 引擎未配置 LLM, 用户上线后请手写回复)`;
     confidence = 0.1;
   }
 

package/src/agents/permission-mode.ts

@@ -0,0 +1,127 @@
+/**
+ * Permission Mode — 3 种模式枚举 + 解析
+ *
+ * 设计: 严格对齐 Claude Code 论文 3 种 mode (default / acceptEdits / bypassPermissions)
+ * 论文第 7 种 (plan/dontAsk/bubble) bolloon 暂不引入, 减少 API surface.
+ *
+ * 关键设计: shell-guard **不受** mode 影响
+ *   - 路径黑名单 (.bolloon/, pi-sdk.ts, shell-guard.ts, .env, .git/ 等) 永远生效
+ *   - 命令 allowlist 默认: git/node/npm/npx/tsx/tsc/vitest/cat/head/tail/wc/ls/echo/pwd/date/mkdir/touch 永远生效
+ *   - 即使用户设 bypassPermissions, shell 类工具仍走 shell-guard
+ *
+ * 失败静默: 任何异常 → 返回 'default' (向后兼容)
+ */
+
+export type PermissionMode = 'default' | 'acceptEdits' | 'bypassPermissions';
+
+export const ALL_PERMISSION_MODES: readonly PermissionMode[] = [
+  'default',
+  'acceptEdits',
+  'bypassPermissions',
+] as const;
+
+export interface ResolveOptions {
+  /** BootstrapOptions / onPreToolUse 调用处透传 */
+  permissionMode?: PermissionMode;
+}
+
+/**
+ * 解析当前 permission mode.
+ * 优先级 (由高到低):
+ *   1. opts.permissionMode          (BootstrapOptions 显式传入)
+ *   2. runtime override             (~/.bolloon/sessions/permission-mode.json, UI 设置)
+ *   3. env BOLLOON_PERM_MODE        (环境变量)
+ *   4. 'default'                    (兜底)
+ *
+ * 非法值 → fallback 'default', console.warn
+ * 失败静默: 任何异常 → 'default' (不阻塞主对话)
+ */
+export function resolvePermissionMode(opts?: ResolveOptions): PermissionMode {
+  try {
+    if (opts?.permissionMode && ALL_PERMISSION_MODES.includes(opts.permissionMode)) {
+      return opts.permissionMode;
+    }
+    // runtime override (UI 设置, 存 ~/.bolloon/sessions/permission-mode.json)
+    const override = readRuntimeOverride();
+    if (override) {
+      return override;
+    }
+    const env = process.env.BOLLOON_PERM_MODE;
+    if (env === 'default' || env === 'acceptEdits' || env === 'bypassPermissions') {
+      return env;
+    }
+    if (env && env.length > 0) {
+      console.warn(`[permission-mode] unknown BOLLOON_PERM_MODE="${env}", fallback to "default"`);
+    }
+    return 'default';
+  } catch (err) {
+    console.warn('[permission-mode] resolvePermissionMode failed (silent, using default):', err);
+    return 'default';
+  }
+}
+
+/**
+ * 读 runtime override 文件 (UI 设的 mode).
+ * 失败静默: 任何异常 → null (回到 env/default)
+ */
+function readRuntimeOverride(): PermissionMode | null {
+  try {
+    const fs = require('fs') as typeof import('fs');
+    const os = require('os') as typeof import('os');
+    const path = require('path') as typeof import('path');
+    const file = path.join(
+      process.env.HOME || os.homedir() || '/tmp',
+      '.bolloon', 'sessions', 'permission-mode.json'
+    );
+    if (!fs.existsSync(file)) return null;
+    const raw = fs.readFileSync(file, 'utf-8');
+    const obj = JSON.parse(raw);
+    if (obj && typeof obj.mode === 'string' && ALL_PERMISSION_MODES.includes(obj.mode as any)) {
+      return obj.mode as PermissionMode;
+    }
+    return null;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * 工具名分类: 是不是"编辑"类 (受 acceptEdits 模式影响)
+ * 简单按前缀, 不做完整注册表
+ */
+export function isEditTool(toolName: string): boolean {
+  const lc = toolName.toLowerCase();
+  return lc.startsWith('edit_') || lc.startsWith('write_') || lc === 'str_replace' || lc === 'create_file' || lc === 'present_files';
+}
+
+/**
+ * 工具名分类: 是不是"shell"类 (永远走 shell-guard, 不受 mode 影响)
+ */
+export function isShellTool(toolName: string): boolean {
+  const lc = toolName.toLowerCase();
+  return lc === 'shell' || lc === 'shell_exec' || lc === 'bash';
+}
+
+/**
+ * 判断当前 mode + tool 的最终是否需要走 blacklist
+ *
+ * 逻辑表:
+ *   - shell 工具: 永远走 blacklist (shell-guard 接管)
+ *   - bypassPermissions 模式: 非 shell 全部放行
+ *   - acceptEdits 模式: edit_* / write_* 类跳过 blacklist
+ *   - default 模式: 全部走 blacklist
+ */
+export function shouldRunBlacklist(toolName: string, mode: PermissionMode): boolean {
+  if (isShellTool(toolName)) return true;  // shell-guard 永远跑
+  if (mode === 'bypassPermissions') return false;
+  if (mode === 'acceptEdits' && isEditTool(toolName)) return false;
+  return true;  // default
+}
+
+// ============================================================
+// 测试钩子
+// ============================================================
+
+export function _resetPermissionModeForTest(): void {
+  // 解析只读 env, 无内部状态, 保留 API 一致
+}