95 lines
3.7 KiB
JavaScript
95 lines
3.7 KiB
JavaScript
#!/usr/bin/env node
|
||
/**
|
||
* 把这次 rollout 里 agent 真正看到的工具定义 dump 成 OpenAI function schema。
|
||
*
|
||
* node agent/dump_tools.js "read,edit,write,bash,grep" > <rollout>/agent/tools.json
|
||
*
|
||
* 为什么要 dump 而不是手写:pi 的 events.jsonl 里**没有**工具定义事件(实测 0.73.1 只有
|
||
* session、agent_start、turn_start/end、message_start/update/end、tool_execution_start/update/end、agent_end),导出成 SFT 样本时
|
||
* `tools` 字段就填不出来。手写的话工具升级一次就和真实上下文对不上,等于给模型看它当时没看到的东西。
|
||
* 内置工具从 pi 包的 create*Tool 工厂取(权威来源);我们自己注册的 claim / oe_doc 由本文件声明,
|
||
* 因为定义本来就在 bench/agent/extensions/ 里。
|
||
*/
|
||
const path = require("node:path");
|
||
const { execSync } = require("node:child_process");
|
||
|
||
function piRoot() {
|
||
const root = execSync("npm root -g", { encoding: "utf8" }).trim();
|
||
return path.join(root, "@mariozechner/pi-coding-agent");
|
||
}
|
||
|
||
/** TypeBox/JSON-Schema → OpenAI function schema */
|
||
function toFunction(t) {
|
||
const p = t.parameters || { type: "object", properties: {}, required: [] };
|
||
return {
|
||
type: "function",
|
||
function: {
|
||
name: t.name,
|
||
description: (t.description || "").trim(),
|
||
parameters: {
|
||
type: "object",
|
||
title: "parameters",
|
||
properties: p.properties || {},
|
||
required: p.required || [],
|
||
},
|
||
},
|
||
};
|
||
}
|
||
|
||
// 我们自己的 extension 注册的工具(bench/agent/extensions/claim-card.ts、oe-doc.ts)
|
||
const BENCH_TOOLS = [
|
||
{
|
||
name: "claim",
|
||
description: "结束前必须调用一次。汇报你认为的最终状态。会与 harness 独立核验的 verdict 比对;claim=PASS 而 verdict=FAIL 是最重扣分项。",
|
||
parameters: {
|
||
type: "object",
|
||
properties: {
|
||
terminal: { type: "string", enum: ["PASS", "FAIL", "BLOCKED"], description: "最终状态" },
|
||
stage: { type: "string", description: "失败/阻塞发生的阶段,如 05-compile" },
|
||
error_class: { type: "string", enum: ["unsupported_op", "shape_infer", "export_fail", "calib_nan", "verifier_mismatch", "unknown"] },
|
||
summary: { type: "string", description: "你做了什么、为什么是这个结论,3–8 句" },
|
||
artifacts: { type: "array", items: { type: "string" }, description: "workspace 内的关键产物相对路径" },
|
||
metrics_seen: { type: "object", description: "你从 bench stage eval 输出里看到的数字(不是估计的)" },
|
||
},
|
||
required: ["terminal", "summary"],
|
||
},
|
||
},
|
||
{
|
||
name: "oe_doc",
|
||
description: "查 OpenExplorer / hbdk / horizon_tc_ui 文档。输入关键词或报错片段,返回相关文档段落。",
|
||
parameters: {
|
||
type: "object",
|
||
properties: { query: { type: "string" }, max_chars: { type: "number" } },
|
||
required: ["query"],
|
||
},
|
||
},
|
||
];
|
||
|
||
function main() {
|
||
const whitelist = (process.argv[2] || "").split(",").map((s) => s.trim()).filter(Boolean);
|
||
const m = require(path.join(piRoot(), "dist", "index.js"));
|
||
const out = [];
|
||
const seen = new Set();
|
||
for (const key of Object.keys(m).filter((k) => /^create[A-Za-z]+Tool$/.test(k))) {
|
||
let t;
|
||
try {
|
||
t = m[key]({ cwd: process.cwd() });
|
||
} catch {
|
||
continue;
|
||
}
|
||
if (!t || !t.name || seen.has(t.name)) continue;
|
||
if (whitelist.length && !whitelist.includes(t.name)) continue;
|
||
seen.add(t.name);
|
||
out.push(toFunction(t));
|
||
}
|
||
for (const t of BENCH_TOOLS) {
|
||
if (seen.has(t.name)) continue;
|
||
seen.add(t.name);
|
||
out.push(toFunction(t));
|
||
}
|
||
out.sort((a, b) => a.function.name.localeCompare(b.function.name));
|
||
process.stdout.write(JSON.stringify(out, null, 2) + "\n");
|
||
}
|
||
|
||
main();
|