braintrust-lite 0.1.4 → 0.1.6

package/bin/consult

@@ -32,7 +32,7 @@ function printHelp() {
        cat file | consult "explain this"
 
 Options:
-  --only <model>     Only run one model: codex | gemini
+  --only <model>     Only run one model: codex | gemini | claude
   --skip <model>     Skip a model (repeatable)
   --timeout <sec>    Per-model timeout in seconds (default: 90)
   --dir <path>       Working directory for CLI subprocesses
@@ -56,12 +56,13 @@ if (!prompt) {
 
 // ─── Run ─────────────────────────────────────────────────────────────────────
 
-const results = await consult({
+const { results, mapping } = await consult({
   prompt,
   only: flags.only,
   skip: flags.skip,
-  timeoutMs: flags.timeout ? flags.timeout * 1000 : 90_000,
+  timeoutMs: flags.timeout != null ? flags.timeout * 1000 : 90_000,
   cwd: flags.dir ? resolve(flags.dir) : undefined,
+  blind: !flags.only, // blind mode only useful when multiple providers run
 });
 
 // Progress summary to stderr
@@ -72,7 +73,7 @@ for (const r of results) {
 
 // Output to stdout
 if (flags.json) {
-  console.log(formatAsJson(prompt, results));
+  console.log(formatAsJson(prompt, results, mapping));
 } else {
-  console.log('\n' + formatAsMarkdown(results));
+  console.log('\n' + formatAsMarkdown(results, mapping));
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "braintrust-lite",
-  "version": "0.1.4",
+  "version": "0.1.6",
   "description": "Lightweight multi-model advisor for Claude Code — parallel Codex + Gemini consultation via MCP",
   "type": "module",
   "bin": {

package/src/consult.js

@@ -43,11 +43,17 @@ function shuffle(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]);
-  return shuffled.map((r, i) => ({ ...r, provider: labels[i] }));
+  const mapping = {};
+  const anonymized = shuffled.map((r, i) => {
+    mapping[labels[i]] = r.provider;
+    return { ...r, provider: labels[i] };
+  });
+  return { results: anonymized, mapping };
 }
 
 /**
@@ -77,8 +83,8 @@ async function runOne(name, prompt, { cwd, timeoutMs }) {
  * @param {string[]} [opts.skip]       - Providers to skip.
  * @param {number}   [opts.timeoutMs]  - Per-provider timeout in ms (default 90 000). 0 = no timeout.
  * @param {string}   [opts.cwd]        - Working directory for subprocesses.
- * @param {boolean}  [opts.blind]      - Anonymise provider names in output (default true).
- * @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, blind = true } = {}) {
   const targets = Object.keys(PROVIDERS)
@@ -99,5 +105,8 @@ export async function consult({ prompt, only, skip = [], timeoutMs = 90_000, cwd
       : { provider: name, content: '', duration_ms: 0, error: settled[i].reason?.message ?? 'unknown' }
   );
 
-  return blind ? anonymize(results) : results;
+  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/server.js

@@ -68,7 +68,7 @@ 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 : [],
@@ -84,7 +84,7 @@ server.setRequestHandler(CallToolRequestSchema, async req => {
   }
 
   return {
-    content: [{ type: 'text', text: formatAsMarkdown(results) }],
+    content: [{ type: 'text', text: formatAsMarkdown(results, mapping) }],
   };
 });