braintrust-lite 0.1.3 → 0.1.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "braintrust-lite",
-  "version": "0.1.3",
+  "version": "0.1.5",
   "description": "Lightweight multi-model advisor for Claude Code — parallel Codex + Gemini consultation via MCP",
   "type": "module",
   "bin": {

package/src/consult.js

@@ -2,8 +2,10 @@ import {
   runProcess,
   adaptCodex,
   adaptGemini,
+  adaptClaude,
   CODEX_ARGS_PREFIX,
   GEMINI_ARGS_PREFIX,
+  CLAUDE_ARGS_PREFIX,
 } from './providers.js';
 
 const SYSTEM_PROMPT = `你是一个独立思考的高级专家。请基于自己的判断给出高质量、可执行的回答。
@@ -20,8 +22,40 @@ const PROVIDERS = {
     buildArgs: prompt => ['-p', `${SYSTEM_PROMPT}\n\n${prompt}`, ...GEMINI_ARGS_PREFIX],
     adapt: adaptGemini,
   },
+  claude: {
+    cmd: 'claude',
+    buildArgs: prompt => [...CLAUDE_ARGS_PREFIX, `${SYSTEM_PROMPT}\n\n${prompt}`],
+    adapt: adaptClaude,
+  },
 };
 
+/**
+ * Shuffle an array in-place using Fisher-Yates and return it.
+ */
+function shuffle(arr) {
+  for (let i = arr.length - 1; i > 0; i--) {
+    const j = Math.floor(Math.random() * (i + 1));
+    [arr[i], arr[j]] = [arr[j], arr[i]];
+  }
+  return arr;
+}
+
+/**
+ * Replace provider names with anonymous labels (Model A, B, C…).
+ * Order is randomised so the judge cannot infer identity from position.
+ * Returns { results: anonymized array, mapping: { 'Model A': 'gemini', … } }
+ */
+function anonymize(results) {
+  const labels = ['Model A', 'Model B', 'Model C', 'Model D', 'Model E'];
+  const shuffled = shuffle([...results]);
+  const mapping = {};
+  const anonymized = shuffled.map((r, i) => {
+    mapping[labels[i]] = r.provider;
+    return { ...r, provider: labels[i] };
+  });
+  return { results: anonymized, mapping };
+}
+
 /**
  * Run a single provider and return a normalized result object.
  * Never throws — errors are captured in the `error` field.
@@ -41,17 +75,18 @@ async function runOne(name, prompt, { cwd, timeoutMs }) {
 }
 
 /**
- * Consult Codex and/or Gemini in parallel.
+ * Consult Codex, Gemini, and/or Claude in parallel.
  *
  * @param {object} opts
  * @param {string}   opts.prompt       - The question to ask.
- * @param {string}   [opts.only]       - 'codex' | 'gemini' — only run this one.
+ * @param {string}   [opts.only]       - 'codex' | 'gemini' | 'claude' — only run this one.
  * @param {string[]} [opts.skip]       - Providers to skip.
- * @param {number}   [opts.timeoutMs]  - Per-provider timeout in ms (default 90 000).
+ * @param {number}   [opts.timeoutMs]  - Per-provider timeout in ms (default 90 000). 0 = no timeout.
  * @param {string}   [opts.cwd]        - Working directory for subprocesses.
- * @returns {Promise<Array<{provider, content, duration_ms, error}>>}
+ * @param {boolean}  [opts.blind]      - Anonymise provider names (default true).
+ * @returns {Promise<{ results: Array, mapping: object|null }>}
  */
-export async function consult({ prompt, only, skip = [], timeoutMs = 90_000, cwd } = {}) {
+export async function consult({ prompt, only, skip = [], timeoutMs = 90_000, cwd, blind = true } = {}) {
   const targets = Object.keys(PROVIDERS)
     .filter(name => (only ? name === only : true))
     .filter(name => !skip.includes(name));
@@ -64,9 +99,14 @@ export async function consult({ prompt, only, skip = [], timeoutMs = 90_000, cwd
     targets.map(name => runOne(name, prompt, { cwd, timeoutMs }))
   );
 
-  return targets.map((name, i) =>
+  const results = targets.map((name, i) =>
     settled[i].status === 'fulfilled'
       ? settled[i].value
       : { provider: name, content: '', duration_ms: 0, error: settled[i].reason?.message ?? 'unknown' }
   );
+
+  if (blind) {
+    return anonymize(results);
+  }
+  return { results, mapping: null };
 }

package/src/format.js

@@ -1,20 +1,42 @@
 /**
  * Format an array of provider results as human-readable Markdown.
  * Each provider gets a ## header with timing (or error), then its content.
+ *
+ * When a mapping is provided (blind mode), a REVEAL section is appended
+ * at the end so the Judge can disclose model identity to the user AFTER
+ * completing their evaluation.
  */
-export function formatAsMarkdown(results) {
-  return results.map(r => {
+export function formatAsMarkdown(results, mapping = null) {
+  const body = results.map(r => {
     const label = r.error
-      ? `## ${r.provider.toUpperCase()} (${r.error})`
-      : `## ${r.provider.toUpperCase()} (${(r.duration_ms / 1000).toFixed(1)}s)`;
-    const body = r.error ? `*调用失败: ${r.error}*` : r.content;
-    return `${label}\n\n${body}`;
+      ? `## ${r.provider} (${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');
+
+  if (!mapping) return body;
+
+  const reveal = Object.entries(mapping)
+    .map(([label, provider]) => `| ${label} | **${provider}** |`)
+    .join('\n');
+
+  return `${body}
+
+---
+
+## 🔒 REVEAL — 仅在完成评估后阅读
+
+> **Judge 指令**：请先完成你的完整评估和综合输出，再阅读以下映射表，并在回复末尾告知用户每个模型对应的真实身份。
+
+| 匿名标签 | 真实模型 |
+|---------|---------|
+${reveal}`;
 }
 
 /**
  * Format results as a compact JSON string for programmatic consumption.
  */
-export function formatAsJson(prompt, results) {
-  return JSON.stringify({ prompt, results }, null, 2);
+export function formatAsJson(prompt, results, mapping = null) {
+  return JSON.stringify({ prompt, results, mapping }, null, 2);
 }

package/src/providers.js

@@ -4,6 +4,7 @@ import { spawn } from 'child_process';
 
 export const CODEX_ARGS_PREFIX = ['exec', '--json', '--skip-git-repo-check', '--ephemeral'];
 export const GEMINI_ARGS_PREFIX = ['-o', 'json'];
+export const CLAUDE_ARGS_PREFIX = ['--output-format', 'json', '-p'];
 
 // ─── Process runner ───────────────────────────────────────────────────────────
 
@@ -76,3 +77,12 @@ export function adaptGemini(raw) {
   } catch { /* fall through */ }
   return fallback(raw.stdout);
 }
+
+/** Parse Claude CLI JSON output → extract .result text. */
+export function adaptClaude(raw) {
+  try {
+    const j = JSON.parse(raw.stdout);
+    if (typeof j.result === 'string') return { content: j.result, parse_mode: 'json' };
+  } catch { /* fall through */ }
+  return fallback(raw.stdout);
+}

package/src/server.js

@@ -26,18 +26,22 @@ const CONSULT_TOOL = {
       },
       only: {
         type: 'string',
-        enum: ['codex', 'gemini'],
+        enum: ['codex', 'gemini', 'claude'],
         description: '只调用指定一个模型（省成本或调试）。',
       },
       skip: {
         type: 'array',
-        items: { type: 'string', enum: ['codex', 'gemini'] },
+        items: { type: 'string', enum: ['codex', 'gemini', 'claude'] },
         description: '跳过指定模型列表。',
       },
       timeout_sec: {
         type: 'number',
         description: '每个模型的超时秒数，默认 90。传 0 表示不限时，等待直到完成。',
       },
+      blind: {
+        type: 'boolean',
+        description: '匿名化 provider 名称（默认 true）。结果以 Model A/B/C 呈现，顺序随机打乱，Judge 无法通过名字或位置判断来源。传 false 可查看真实模型名称。',
+      },
       cwd: {
         type: 'string',
         description: '子进程工作目录，默认继承 MCP server 的 cwd。',
@@ -64,11 +68,12 @@ server.setRequestHandler(CallToolRequestSchema, async req => {
   }
 
   const args = req.params.arguments ?? {};
-  const results = await consult({
+  const { results, mapping } = await consult({
     prompt: String(args.prompt ?? ''),
     only: args.only,
     skip: Array.isArray(args.skip) ? args.skip : [],
     timeoutMs: args.timeout_sec != null ? Number(args.timeout_sec) * 1000 : 90_000,
+    blind: args.blind !== false,
     cwd: args.cwd,
   });
 
@@ -79,7 +84,7 @@ server.setRequestHandler(CallToolRequestSchema, async req => {
   }
 
   return {
-    content: [{ type: 'text', text: formatAsMarkdown(results) }],
+    content: [{ type: 'text', text: formatAsMarkdown(results, mapping) }],
   };
 });