braintrust-lite 0.1.7 → 0.1.8

package/README.md

@@ -1,149 +1,120 @@
-# braintrust-lite
+# brantrust
 
-Claude Code 原生的多模型军师 — 并发调用 Codex + Gemini，主 Claude 担任 Judge 融合输出。
+同题多模型融合器 — 把同一个问题同时发给 Claude、Codex、Gemini，然后用一个 Judge 综合出"集大成方案"。
 
 ```
-主 Claude → parallel:
-  ├─ Task(subagent_type=Plan, prompt=X)       ← 正常子 agent
-  └─ mcp__braintrust_lite__consult(prompt=X) ← Codex + Gemini 旁路咨询
-          → 主 Claude 融合三方视角 → 最终方案
+输入 → 并发生成(3) → 清洗归一化 → Judge 融合(1) → 输出 + 落盘
 ```
 
-vs [`braintrust`](https://github.com/HongjieRen/braintrust): 2 次 API 调用（省 50%），无独立 Judge，无落盘，原生集成 Claude Code。
+4 次 API 调用，低成本，天天能用。
 
 ---
 
 ## 安装
 
-**前置条件**：`codex` 和 `gemini` CLI 均已登录。
-
 ```bash
 # 克隆
-git clone https://github.com/HongjieRen/braintrust-lite.git
-cd braintrust-lite
-
-# 安装依赖
-npm install
-
-# 可选：把 CLI 软链到 PATH
-ln -sf "$(pwd)/bin/consult" ~/.local/bin/consult
-chmod +x bin/consult
-```
+git clone https://github.com/HongjieRen/brantrust.git
+cd brantrust
 
----
-
-## 注册到 Claude Code（MCP）
-
-```bash
-claude mcp add braintrust-lite node "$(pwd)/src/server.js"
+# 软链接到 PATH
+ln -sf "$(pwd)/brantrust" ~/.local/bin/brantrust
+chmod +x brantrust
 ```
 
-注册后，Claude Code 会话里会出现 `mcp__braintrust_lite__consult` tool，和 `Read` / `Bash` 并列可用。
-
-重启 Claude Code 后生效。
-
----
-
-## 安装 Skill 引导
-
-把 skill 软链到 Claude Code 全局 skill 目录，让主 Claude 知道何时该主动使用 consult：
-
-```bash
-ln -sf "$(pwd)/skills/consult" ~/.claude/skills/consult
-```
+**前置依赖**（三个 CLI 均需已登录）：
 
-安装后可用 `/consult` slash command 激活"军师模式"引导。
+| Provider | CLI | 验证命令 |
+|----------|-----|---------|
+| Claude | `claude` | `claude -p "hi" --output-format json` |
+| OpenAI Codex | `codex` | `codex exec "hi" --json --skip-git-repo-check --ephemeral` |
+| Google Gemini | `gemini` | `gemini -p "hi" -o json` |
 
 ---
 
-## 使用方式
-
-### 在 Claude Code 里（推荐）
-
-Claude 会在处理规划/设计类任务时自动（或在 `/consult` 引导下）并发调用：
-
-```
-你处理一个架构选型任务时，Claude 会同时：
-  1. 启动 Plan sub-agent 做深度分析
-  2. 调用 mcp__braintrust_lite__consult 获取 Codex + Gemini 的独立视角
-  3. 融合三方输出给你最终方案
-```
-
-### 终端 CLI（fallback / 调试）
+## 用法
 
 ```bash
-consult "解释 CAP 定理"                   # 并发两模型，markdown 输出
-consult --only codex "prompt"             # 只跑 codex
-consult --skip gemini "prompt"            # 跳过 gemini
-consult --timeout 60 "prompt"             # 超时秒数
-consult --dir ~/myproject "review"        # 工作目录
-cat app.ts | consult "review this code"   # stdin 拼接
-consult --json "prompt"                   # JSON 结构化输出
+brantrust "解释 CAP 定理"                      # 默认：3 generator + 1 judge
+brantrust --no-judge "React vs Vue"            # 只并发收集，不 judge
+brantrust --judge-model gemini "数据库选型"    # 切换 Judge 模型
+brantrust --skip codex "量子计算"              # 跳过某个模型（可多次）
+cat app.ts | brantrust "review 这段代码"       # stdin 管道
+brantrust --dir ~/project "项目分析"           # 指定工作目录
+brantrust --context-file design.md "实现方案"  # 附加上下文文件
+brantrust --timeout 60 "快速问题"              # 超时秒数
+brantrust --no-save "临时问答"                 # 不保存到磁盘
+brantrust --json "问题"                        # 输出完整 JSON
+brantrust --list                               # 查看历史运行
+brantrust --strict "关键决策"                  # [v2] 完整 Judge 流水线
 ```
 
----
-
-## 参数
+### 参数一览
 
 | 参数 | 默认 | 说明 |
-|---|---|---|
-| `prompt` | 必须 | 问题文本（MCP）/ 位置参数（CLI）|
-| `only` | — | 只调用: `codex` \| `gemini` |
-| `skip` | — | 跳过模型列表 |
-| `timeout_sec` | `90` | 每个模型超时秒数 |
-| `cwd` | server cwd | 子进程工作目录 |
-| `--json` | false | CLI 专用：JSON 格式输出 |
+|------|------|------|
+| `"prompt"` | 必须 | 问题文本 |
+| `--skip <model>` | — | 跳过模型：claude / codex / gemini，可多次使用 |
+| `--judge-model <model>` | `claude` | Judge 使用的模型 |
+| `--no-judge` | false | 关闭 Judge，只展示各模型原始回答 |
+| `--timeout <sec>` | `120` | 每个模型的超时秒数 |
+| `--dir <path>` | cwd | CLI 工具的工作目录 |
+| `--context-file <file>` | — | 附加文件内容作为上下文（最多 8000 字符）|
+| `--no-save` | false | 不保存结果到磁盘 |
+| `--json` | false | 将完整结果以 JSON 格式输出到 stdout |
+| `--list` | — | 列出最近 20 条历史运行 |
+| `--strict` | — | [v2 占位] 两阶段 Judge + swap-compare |
 
 ---
 
-## 输出格式
+## 输出
 
-```
-## CODEX (8.2s)
-
-<codex 完整回答>
-
----
+**终端**：各模型回答 + Judge 融合报告（Markdown 格式）
 
-## GEMINI (6.5s)
+**落盘**（`~/ai-outputs/<timestamp>/`）：
 
-<gemini 完整回答>
 ```
-
-失败的 provider 显示 `*调用失败: timeout*`，另一个照常返回（`Promise.allSettled` 容错）。
+~/ai-outputs/2026-04-09T11-23-45-678/
+├── raw/
+│   ├── claude.txt
+│   ├── codex.txt
+│   └── gemini.txt
+├── normalized.json    # 三个模型的结构化摘要
+└── report.md          # 最终融合报告
+```
 
 ---
 
 ## 架构
 
 ```
-braintrust-lite/
-├── src/
-│   ├── server.js      MCP stdio server
-│   ├── consult.js     核心并发逻辑
-│   ├── providers.js   spawn + Codex/Gemini 解析器
-│   └── format.js      Markdown / JSON 渲染
-├── bin/
-│   └── consult        CLI 入口
-├── skills/
-│   └── consult/
-│       └── SKILL.md   Claude Code skill 引导
-└── docs/
-    └── spec.md        设计文档
+runGenerators()         # 并发调用三个 CLI，AbortController 超时，Promise.allSettled 容错
+normalizeResults()      # 各适配器提取 content / key_claims / assumptions / risks
+runSimpleJudge()        # 单次 Judge 调用，只传归一化摘要（非全文），控制 token
+writeRunArtifacts()     # 落盘 raw/ + normalized.json + report.md
+runFullJudgePipeline()  # [v2 占位] 两阶段 Judge + swap-compare + 抗偏置
 ```
 
+**Judge prompt 匿名化**：候选标签只用 A / B / C，不暴露 provider 名称，避免模型偏置。
+
 ---
 
-## 成本
+## 成本估算
+
+每次运行 = 4 次 API 调用（3 generator + 1 judge）：
 
-| 场景 | API 调用 | 估算成本 |
-|---|---|---|
-| 简单问题 | 2 | $0.05–0.15 |
-| 中等问题 | 2 | $0.15–0.40 |
-| 复杂问题 | 2 | $0.40–0.80 |
+| 问题复杂度 | 估算成本 |
+|-----------|---------|
+| 简单 | $0.20 – 0.50 |
+| 中等 | $0.50 – 1.00 |
+| 复杂 | $1.00 – 2.00 |
 
 ---
 
-## License
+## V2 路线图
 
-MIT
+1. `--strict`：两阶段 Judge (A+B) + swap-compare + 抗偏置
+2. `--continue`：线程续聊
+3. `--context-file` 智能截断 + git diff 注入
+4. 成本 / token 预算控制器
+5. 更多 provider（Goose、本地模型等）

package/bin/braintrust

@@ -0,0 +1,12 @@
+#!/usr/bin/env node
+'use strict';
+
+// Shim: delegates to src/main.js
+// The symlink ~/.local/bin/braintrust → this file remains unchanged.
+
+const { main } = require('../src/main.js');
+
+main(process.argv.slice(2)).catch(e => {
+  process.stderr.write(`[braintrust error] ${e.message}\n`);
+  process.exit(1);
+});

package/package.json

@@ -1,40 +1,40 @@
 {
   "name": "braintrust-lite",
-  "version": "0.1.7",
-  "description": "Lightweight multi-model advisor for Claude Code — parallel Codex + Gemini consultation via MCP",
-  "type": "module",
+  "version": "0.1.8",
+  "description": "Multi-model AI consultation MCP for Claude Code — runs Claude, Codex, and Gemini in parallel for Judge-style synthesis",
   "bin": {
-    "consult": "bin/consult",
-    "braintrust-lite": "src/server.js",
-    "braintrust-setup": "scripts/setup.js"
+    "braintrust-lite": "./src/server.js",
+    "braintrust": "./bin/braintrust",
+    "braintrust-doctor": "./src/doctor.js"
   },
+  "main": "src/main.js",
+  "files": [
+    "src/",
+    "skills/",
+    "bin/",
+    "README.md"
+  ],
   "scripts": {
-    "start": "node src/server.js",
-    "setup": "node scripts/setup.js"
+    "test": "node --test src/normalize.test.js"
   },
   "dependencies": {
-    "@modelcontextprotocol/sdk": "^1.10.2"
+    "better-sqlite3": "^11.0.0"
   },
   "engines": {
-    "node": ">=18"
+    "node": ">=18.0.0"
   },
   "keywords": [
     "mcp",
     "claude-code",
+    "multi-model",
+    "ai",
     "codex",
     "gemini",
-    "multi-model",
-    "ai"
+    "braintrust"
   ],
-  "license": "MIT",
   "repository": {
     "type": "git",
-    "url": "git+https://github.com/HongjieRen/braintrust-lite.git"
+    "url": "https://github.com/HongjieRen/braintrust.git"
   },
-  "files": [
-    "bin/",
-    "src/",
-    "scripts/",
-    "skills/"
-  ]
+  "license": "MIT"
 }

package/skills/consult/SKILL.md

@@ -9,7 +9,7 @@ description: 在处理规划/设计/架构/调研类任务时，并发调用 cod
 ## Preamble（每次调用前先执行）
 
 ```bash
-_REMOTE_VER=$(curl -sf --max-time 3 "https://raw.githubusercontent.com/HongjieRen/braintrust-lite/main/skills/consult/SKILL.md" | awk '/^version:/ {print $2; exit}')
+_REMOTE_VER=$(curl -sf --max-time 3 "https://raw.githubusercontent.com/HongjieRen/braintrust/main/skills/consult/SKILL.md" | awk '/^version:/ {print $2; exit}')
 _LOCAL_VER=$(awk '/^version:/ {print $2; exit}' ~/.claude/skills/consult/SKILL.md 2>/dev/null)
 [ -n "$_REMOTE_VER" ] && [ "$_REMOTE_VER" != "$_LOCAL_VER" ] && echo "UPGRADE_AVAILABLE $_LOCAL_VER $_REMOTE_VER" || echo "UP_TO_DATE $_LOCAL_VER"
 ```
@@ -17,7 +17,7 @@ _LOCAL_VER=$(awk '/^version:/ {print $2; exit}' ~/.claude/skills/consult/SKILL.m
 **如果输出 `UPGRADE_AVAILABLE <old> <new>`**：静默自动更新，不打断用户：
 ```bash
 cp ~/.claude/skills/consult/SKILL.md ~/.claude/skills/consult/SKILL.md.bak 2>/dev/null || true
-curl -fsSL "https://raw.githubusercontent.com/HongjieRen/braintrust-lite/main/skills/consult/SKILL.md" \
+curl -fsSL "https://raw.githubusercontent.com/HongjieRen/braintrust/main/skills/consult/SKILL.md" \
   -o ~/.claude/skills/consult/SKILL.md && echo "Updated consult skill $_old → $_new"
 ```
 更新完成后继续执行本次任务，在最终回复末尾附一行：`*(consult skill 已自动更新 v{old} → v{new})*`

package/src/config.js

@@ -0,0 +1,60 @@
+'use strict';
+
+const { join } = require('path');
+
+const PROJECT_ROOT = join(__dirname, '..');
+const OUTPUT_DIR = join(PROJECT_ROOT, 'ai-outputs');
+const STATE_DIR = join(OUTPUT_DIR, '.state');
+const DB_PATH = join(STATE_DIR, 'braintrust.sqlite');
+const POLICY_PATH = join(STATE_DIR, 'policy.json');
+const REFLECTOR_LOG = join(STATE_DIR, 'reflector.log');
+
+const DEFAULT_TIMEOUT_S = 120;
+const DEFAULT_JUDGE_MODEL = 'claude';
+const DEFAULT_MEMORY_K = 3;
+const MAX_CONTEXT_CHARS = 30000;
+const CONTEXT_FILE_MAX = 8000;
+
+// Memory injection hard limits (chars)
+const MEMORY_INJECT_LIMIT = 1500;
+const LESSONS_INJECT_LIMIT = 600;
+const SKILLS_INJECT_LIMIT = 800;
+
+// Novelty check threshold: cosine similarity above this → prompt reuse
+const NOVELTY_THRESHOLD = 0.9;
+
+// Critique-revise disagreement threshold
+const DISAGREE_THRESHOLD = 0.5;
+
+// Economy mode: disable all extra LLM calls
+const ECONOMY = process.env.BRAINTRUST_ECONOMY === '1';
+
+// Reflector model: codex with gpt-5.4-mini.
+// Chosen over haiku/flash for better Chinese text quality.
+// Must differ from the default judge model (claude) to avoid self-evaluation bias.
+const REFLECTOR_MODEL = 'gpt-5.4-mini';
+const REFLECTOR_CMD = 'codex';
+const REFLECTOR_ARGS_PREFIX = ['exec', '--json', '--skip-git-repo-check', '--ephemeral', '-m', REFLECTOR_MODEL];
+
+module.exports = {
+  PROJECT_ROOT,
+  OUTPUT_DIR,
+  STATE_DIR,
+  DB_PATH,
+  POLICY_PATH,
+  REFLECTOR_LOG,
+  DEFAULT_TIMEOUT_S,
+  DEFAULT_JUDGE_MODEL,
+  DEFAULT_MEMORY_K,
+  MAX_CONTEXT_CHARS,
+  CONTEXT_FILE_MAX,
+  MEMORY_INJECT_LIMIT,
+  LESSONS_INJECT_LIMIT,
+  SKILLS_INJECT_LIMIT,
+  NOVELTY_THRESHOLD,
+  DISAGREE_THRESHOLD,
+  ECONOMY,
+  REFLECTOR_MODEL,
+  REFLECTOR_CMD,
+  REFLECTOR_ARGS_PREFIX,
+};

package/src/doctor.js

@@ -0,0 +1,120 @@
+#!/usr/bin/env node
+'use strict';
+
+const { execFileSync, spawnSync } = require('child_process');
+const { existsSync, readFileSync } = require('fs');
+const { join } = require('path');
+const { version: PKG_VERSION } = require('../package.json');
+
+const GREEN = '\x1b[32m✓\x1b[0m';
+const RED   = '\x1b[31m✗\x1b[0m';
+const WARN  = '\x1b[33m!\x1b[0m';
+
+function check(label, ok, detail) {
+  const icon = ok === true ? GREEN : ok === 'warn' ? WARN : RED;
+  const line = `  ${icon}  ${label.padEnd(28)} ${detail || ''}`;
+  console.log(line);
+  return ok === true;
+}
+
+function getVersion(cmd, args) {
+  try {
+    const result = spawnSync(cmd, args, { timeout: 5000, encoding: 'utf8' });
+    if (result.status === 0) {
+      return (result.stdout || result.stderr || '').split('\n')[0].trim().slice(0, 40);
+    }
+    return null;
+  } catch {
+    return null;
+  }
+}
+
+function getSkillVersion(skillPath) {
+  try {
+    const content = readFileSync(skillPath, 'utf8');
+    const m = content.match(/^version:\s*(.+)$/m);
+    return m ? m[1].trim() : 'unknown';
+  } catch {
+    return null;
+  }
+}
+
+function checkMcpServer() {
+  // Probe MCP server: send initialize, expect a valid JSON-RPC response
+  const serverPath = join(__dirname, 'server.js');
+  if (!existsSync(serverPath)) return { ok: false, detail: 'src/server.js not found' };
+
+  try {
+    const msg = JSON.stringify({
+      jsonrpc: '2.0', id: 1, method: 'initialize',
+      params: { protocolVersion: '2024-11-05', capabilities: {}, clientInfo: { name: 'doctor', version: '0' } },
+    });
+    const result = spawnSync(process.execPath, [serverPath], {
+      input: msg + '\n',
+      timeout: 5000,
+      encoding: 'utf8',
+    });
+    const line = (result.stdout || '').split('\n').find(l => l.trim().startsWith('{'));
+    if (!line) return { ok: false, detail: 'no JSON response from server' };
+    const resp = JSON.parse(line);
+    if (resp.result && resp.result.serverInfo) {
+      return { ok: true, detail: `v${resp.result.serverInfo.version}` };
+    }
+    return { ok: false, detail: 'unexpected response shape' };
+  } catch (err) {
+    return { ok: false, detail: err.message.slice(0, 60) };
+  }
+}
+
+function main() {
+  console.log(`\nbraintrust doctor  (package v${PKG_VERSION})\n`);
+
+  let allOk = true;
+
+  // ── CLI tools ──────────────────────────────────────────────────────────────
+  console.log('CLI tools:');
+  for (const [cmd, vArgs, installHint] of [
+    ['claude', ['--version'],      'https://claude.ai/download'],
+    ['codex',  ['--version'],      'npm i -g @openai/codex'],
+    ['gemini', ['--version'],      'npm i -g @google/gemini-cli'],
+  ]) {
+    const ver = getVersion(cmd, vArgs);
+    if (ver) {
+      check(cmd, true, ver);
+    } else {
+      check(cmd, false, `not found — ${installHint}`);
+      allOk = false;
+    }
+  }
+
+  // ── MCP server ─────────────────────────────────────────────────────────────
+  console.log('\nMCP server:');
+  const mcp = checkMcpServer();
+  if (!check('braintrust-lite server', mcp.ok, mcp.detail)) allOk = false;
+
+  // ── Skill ──────────────────────────────────────────────────────────────────
+  console.log('\nConsult skill:');
+  const skillPath = join(process.env.HOME || '~', '.claude', 'skills', 'consult', 'SKILL.md');
+  const skillVer = getSkillVersion(skillPath);
+  if (skillVer) {
+    check('SKILL.md installed', true, `v${skillVer}  at ${skillPath}`);
+  } else {
+    check('SKILL.md installed', false, `not found at ${skillPath}`);
+    allOk = false;
+  }
+
+  const bakPath = skillPath + '.bak';
+  check('SKILL.md.bak exists', existsSync(bakPath) ? 'warn' : 'warn',
+    existsSync(bakPath) ? 'backup present' : 'no backup yet (created on first auto-update)');
+
+  // ── Summary ────────────────────────────────────────────────────────────────
+  console.log();
+  if (allOk) {
+    console.log('  \x1b[32mAll checks passed — braintrust is ready.\x1b[0m\n');
+  } else {
+    console.log('  \x1b[31mSome checks failed — fix the issues above before using braintrust.\x1b[0m\n');
+    process.exit(1);
+  }
+}
+
+main();

package/src/format.js

@@ -1,53 +1,30 @@
+'use strict';
+
 /**
- * Format provider results as human-readable Markdown with run manifest.
+ * Format a CLI run manifest summary for terminal output.
+ *
+ * @param {{ results: Array, ts: string, judgeModel: string|null, runDir: string }} opts
+ * @returns {string}
  */
-export function formatAsMarkdown(results, mapping = null, { successCount, totalCount } = {}) {
-  const total = totalCount ?? results.length;
-  const succeeded = successCount ?? results.filter(r => !r.error).length;
-  const degraded = succeeded < total;
-
-  // Status line (mirrors SKILL.md status bar format)
-  const modelsLabel = degraded ? `⚠ ${succeeded}/${total} models` : `${total} models`;
-  const statusLine = `[Consult | ${modelsLabel} | responses below]\n`;
-
-  const body = results.map(r => {
-    const label = r.error
-      ? `## ${r.provider} (${r.error})`
-      : `## ${r.provider} (${(r.duration_ms / 1000).toFixed(1)}s)`;
-    const content = r.error ? `*调用失败: ${r.error}*` : r.content;
-    return `${label}\n\n${content}`;
-  }).join('\n\n---\n\n');
-
-  const revealSection = mapping ? buildReveal(mapping) : '';
-  const manifest = buildManifest(results, { successCount: succeeded, totalCount: total });
-
-  return `${statusLine}\n${body}${revealSection}\n\n---\n\n${manifest}`;
-}
-
-function buildReveal(mapping) {
-  const rows = Object.entries(mapping)
-    .map(([label, provider]) => `| ${label} | **${provider}** |`)
-    .join('\n');
-  return `\n\n---\n\n## 🔒 REVEAL — 仅在完成评估后阅读
-
-> **Judge 指令**：请先完成你的完整评估和综合输出，再阅读以下映射表，并在回复末尾告知用户每个模型对应的真实身份。
-
-| 匿名标签 | 真实模型 |
-|---------|---------|
-${rows}`;
+function formatManifest({ results, ts, judgeModel, runDir }) {
+  const lines = [
+    '## Run Manifest',
+    '',
+    `Timestamp : ${ts}`,
+    `Judge     : ${judgeModel || 'none (--no-judge)'}`,
+    `Saved to  : ${runDir}`,
+    '',
+    'Providers:',
+  ];
+
+  for (const r of results) {
+    const status = r.error
+      ? `✗  ${(r.error_type || r.error).padEnd(12)}`
+      : `✓  ${(r.duration_ms / 1000).toFixed(1)}s  parse_score=${r.parse_score.toFixed(2)}`;
+    lines.push(`  ${r.provider.padEnd(10)} ${status}`);
+  }
+
+  return lines.join('\n');
 }
 
-function buildManifest(results, { successCount, totalCount }) {
-  const ts = new Date().toISOString().slice(0, 19) + 'Z';
-  const degraded = successCount < totalCount;
-  const lines = results.map(r =>
-    r.error
-      ? `  - ${r.provider}: ${r.error_type || r.error}`
-      : `  - ${r.provider}: ${(r.duration_ms / 1000).toFixed(1)}s`
-  ).join('\n');
-  return `**Run manifest** · \`${ts}\` · ${successCount}/${totalCount} models${degraded ? ' ⚠ degraded' : ''}\n${lines}`;
-}
-
-export function formatAsJson(prompt, results, mapping = null) {
-  return JSON.stringify({ prompt, results, mapping }, null, 2);
-}
+module.exports = { formatManifest };

package/src/judge.js

@@ -0,0 +1,87 @@
+'use strict';
+
+const { PROVIDERS } = require('./providers/index.js');
+const { summarize } = require('./normalize.js');
+const { LESSONS_INJECT_LIMIT } = require('./config.js');
+
+/**
+ * Build the judge prompt, optionally injecting lessons from memory.
+ * @param {string} question
+ * @param {Array} results - Normalized provider results
+ * @param {{ lessons?: string[], skills?: string[] }} opts
+ * @returns {string}
+ */
+function buildJudgePrompt(question, results, opts = {}) {
+  const valid = results.filter(r => !r.error);
+  const summaries = valid
+    .map((r, i) => `--- 候选 ${String.fromCharCode(65 + i)} (${r.provider}) ---\n${summarize(r)}`)
+    .join('\n\n');
+
+  const lessonsBlock = buildLessonsBlock(opts.lessons || []);
+
+  return `你是一个高级技术评审。${valid.length} 个 AI 模型对同一问题给出了各自的回答。
+${lessonsBlock}
+问题：${question}
+
+${summaries}
+
+请按以下结构输出你的评审（用中文标签分隔）：
+
+## 核心共识
+（各模型都认同的关键结论）
+
+## 独特洞见
+（某个模型独有但有价值的见解，注明来自哪个候选）
+
+## 分歧裁决
+（如果存在矛盾，给出你的判断和理由；如无分歧则写"无明显分歧"）
+
+## 集大成方案
+（综合各方的最优可执行方案）
+
+## 风险提示
+（需要注意的假设、风险或待验证项）`;
+}
+
+/**
+ * Build a lessons injection block, respecting the hard char limit.
+ * @param {string[]} lessons
+ * @returns {string}
+ */
+function buildLessonsBlock(lessons) {
+  if (!lessons.length) return '';
+  const joined = lessons.slice(0, 5).join('\n');
+  const trimmed = joined.slice(0, LESSONS_INJECT_LIMIT);
+  return `\n<past-lessons>\n${trimmed}\n</past-lessons>\n`;
+}
+
+/**
+ * Run the judge model and return the report text.
+ * @param {string} question
+ * @param {Array} results - Normalized provider results
+ * @param {object} opts
+ * @param {string} [opts.judgeModel='claude'] - Which model to use as judge
+ * @param {Function} opts.runProcess - The process runner function
+ * @param {string[]} [opts.lessons] - Lessons to inject
+ * @returns {Promise<string>}
+ */
+async function runJudge(question, results, opts = {}) {
+  const { judgeModel = 'claude', runProcess, lessons = [] } = opts;
+  const judgePrompt = buildJudgePrompt(question, results, { lessons });
+
+  process.stderr.write(`\n[Judge (${judgeModel}): running...]\n`);
+  const start = Date.now();
+
+  const provider = PROVIDERS[judgeModel];
+  if (!provider) {
+    throw new Error(`Unknown judge model: ${judgeModel}. Use claude|codex|gemini.`);
+  }
+
+  const raw = await runProcess(provider.cmd, provider.getArgs(judgePrompt));
+  const ms = Date.now() - start;
+  process.stderr.write(`[Judge: done ${(ms / 1000).toFixed(1)}s]\n`);
+
+  return provider.extractJudgeText(raw);
+}
+
+module.exports = { buildJudgePrompt, runJudge, buildLessonsBlock };