宜博 / lemonboss

/**
 * Goal Contract — 任务级验收契约的格式、校验、执行与判决持久化（P3 阶段 0 / T0）
 *
 * 设计见 docs/todo/system-design/27-goal-contract-format.md，
 * 需求见 27-p3-p4-work-graph-console.md §五 阶段 0 / §四.A.5 / §八.9，
 * 以及 27-e2e-task-execution-flow.md §十一（三个验收产出节点）。
 *
 * 形态直接借 DevLoop 已跑通的 `integration/spec.json`（_spec/sdd-model.md M-6）
 * 与 `runs/result.json`（M-7/M-8）：契约是一份 JSON、判决是一份 JSON，
 * 都放在约定好的固定位置，谁都不用猜。
 *
 * §八.9 划的优先级：**外层（存哪、怎么被引用、怎么被执行、判决存哪）要定死，
 * 内层断言字段允许后续演进**。所以本文件的稳定面是：
 *
 *   data/companies/{cid}/runs/{rootRunId}/
 *   ├── artifacts/
 *   │   ├── acceptance.json          ← 唯一契约入口（三 section，各自可为 null）
 *   │   ├── goal-benchmark.json      ← 大块内容外置，契约里用 {"$ref": "..."} 指过来
 *   │   ├── goal-test-data/
 *   │   ├── goal-test-script/
 *   │   └── verdicts/
 *   │       ├── {verdictId}.json     ← 一次执行一份完整判决
 *   │       └── index.ndjson         ← append-only 索引，可查、可追溯到 run/node/attempt
 *   └── nodes/{nodeRunId}/           ← output_ref 指向这里，即「被测产物」
 *
 * 与 output_ref 的关联方式：执行器的入参 targetPath 就是某个 graph_run_nodes.output_ref
 * 解析出的绝对路径；判决里回存 run_id / node_run_id / attempt / contract_hash 四元组，
 * 所以任何一份判决都能双向追溯到「哪次 run 的哪个节点的第几次尝试、用哪版契约判的」。
 *
 * 不起 LLM：benchmark / test_data 是纯声明式比对，test_script 走真实子进程。
 *
 * ── 生产调用方（R5-F3）─────────────────────────────────────────────
 * 契约的产出方和消费方都是 **agent**，不是引擎——见 27-goal-contract-format.md §二bis
 * 「为什么接在 agent 工具层而不是引擎钩子」。所以本文件除了作为模块被 require，
 * 还带一个 CLI 入口（文件末尾），供 agent 在 bash 里直接调：
 *
 *   node server/services/goal-contract.js put      --run-dir R --goal-id G --section benchmark --file X.json
 *   node server/services/goal-contract.js validate --run-dir R
 *   node server/services/goal-contract.js run      --run-dir R --target T
 *   node server/services/goal-contract.js list|show --run-dir R [...]
 *
 * 走 CLI 而不是让 agent 自己写 require 脚本，是因为 agent 侧唯一稳定的执行面就是 bash
 * （模板 tools.md 全是 bash 片段），且 `put --file` 让「绝不在 bash 里内联 JSON」这条
 * 铁律得以遵守。
 */

const fs = require('fs');
const path = require('path');
const crypto = require('crypto');
const { spawn } = require('child_process');

// contract_hash 直接复用 DevLoop 的实现（只读借用，不改它一行；见 §七「明确不改」）
const { hashContract } = require('./devloop/doc-store');

const DATA_DIR = path.resolve(__dirname, '../../data');

const CONTRACT_VERSION = '1';
const CONTRACT_FILE = 'acceptance.json';
const SECTIONS = ['benchmark', 'test_data', 'test_script'];
// ponytail: v1 只支持 runner='node'（脚本退出码即判决），够表达一切断言。
// 上限：跑不了 vitest——实测仓库根 vitest.config 的 include 不覆盖 data/，
// `vitest run <脚本>` / `--root` / `--dir` 都会被根 config 的 include 盖掉，直接 "No test files found"。
// 升级路径：真需要 vitest 时，为验收脚本生成一份临时 vitest.config 并用 `--config` 指过去，
// 再把 'vitest' 加回这张白名单——契约的 runner 字段就是为此留的扩展位。
// 眼下 node 脚本里自己 spawn 任何测试框架也不受限制，所以不急。
const RUNNERS = ['node'];
const MATCH_MODES = ['deep-equal', 'subset'];

const DEFAULT_SCRIPT_TIMEOUT_MS = 120_000;
const OUTPUT_CAP = 64 * 1024; // stdout/stderr 各留 64KB，超出截断

class ContractValidationError extends Error {
  constructor(errors) {
    super('goal contract validation failed');
    this.name = 'ContractValidationError';
    this.errors = errors;
  }
}

// ────────────────────────────────────────────────────────────────
// 路径
// ────────────────────────────────────────────────────────────────

/** root run 的产物根目录：data/companies/{cid}/runs/{rootRunId}/ */
function resolveRunDir(companyId, rootRunId) {
  return path.join(DATA_DIR, 'companies', String(companyId), 'runs', String(rootRunId));
}

/** 任务级共享产物区（跨 org node 只读，写入是特权动作，见 §四.A.5） */
function resolveArtifactsDir(runDir) {
  return path.join(runDir, 'artifacts');
}

function resolveContractPath(runDir) {
  return path.join(resolveArtifactsDir(runDir), CONTRACT_FILE);
}

function resolveVerdictsDir(runDir) {
  return path.join(resolveArtifactsDir(runDir), 'verdicts');
}

/** 契约里的相对路径一律相对 artifacts/，且不许逃出 artifacts/（防契约指向任意可执行文件） */
function resolveInsideArtifacts(artifactsDir, rel) {
  const abs = path.resolve(artifactsDir, rel);
  const base = path.resolve(artifactsDir);
  if (abs !== base && !abs.startsWith(base + path.sep)) return null;
  return abs;
}

// ────────────────────────────────────────────────────────────────
// 小工具
// ────────────────────────────────────────────────────────────────

function now() {
  return new Date().toISOString();
}

/** CONVENTIONS §2 的 ID 形态：vd-{YYYYMMDD-HHMM}-{4hex} */
function genVerdictId(d = new Date()) {
  const p2 = (n) => String(n).padStart(2, '0');
  const stamp =
    `${d.getFullYear()}${p2(d.getMonth() + 1)}${p2(d.getDate())}` +
    `-${p2(d.getHours())}${p2(d.getMinutes())}`;
  return `vd-${stamp}-${crypto.randomBytes(2).toString('hex')}`;
}

/** key 排序的规范序列化，用于 deep-equal 比较 */
function stable(v) {
  if (v === null || typeof v !== 'object') return JSON.stringify(v) ?? 'null';
  if (Array.isArray(v)) return '[' + v.map(stable).join(',') + ']';
  return (
    '{' +
    Object.keys(v)
      .sort()
      .map((k) => JSON.stringify(k) + ':' + stable(v[k]))
      .join(',') +
    '}'
  );
}

function deepEqual(a, b) {
  return stable(a) === stable(b);
}

/** subset：expected 里出现的键/元素必须匹配，actual 允许有额外键 */
function matchSubset(expected, actual) {
  if (expected === null || typeof expected !== 'object') return expected === actual;
  if (Array.isArray(expected)) {
    if (!Array.isArray(actual) || actual.length !== expected.length) return false;
    return expected.every((e, i) => matchSubset(e, actual[i]));
  }
  if (actual === null || typeof actual !== 'object' || Array.isArray(actual)) return false;
  return Object.keys(expected).every((k) => matchSubset(expected[k], actual[k]));
}

function readJsonFile(abs) {
  try {
    return { ok: true, json: JSON.parse(fs.readFileSync(abs, 'utf8')) };
  } catch (e) {
    return { ok: false, error: e.message };
  }
}

/** CONVENTIONS §5：JSON 落盘走 tmp + rename */
function writeJsonAtomic(abs, json) {
  fs.mkdirSync(path.dirname(abs), { recursive: true });
  const tmp = abs + '.tmp';
  fs.writeFileSync(tmp, JSON.stringify(json, null, 2));
  fs.renameSync(tmp, abs);
  return abs;
}

// ────────────────────────────────────────────────────────────────
// 契约：读 / 写 / 校验
// ────────────────────────────────────────────────────────────────

/**
 * 校验契约本身是否合法（不碰被测产物）。
 *
 * errors 形态与 org-graph-service.validateGraph 保持一致：{rule, detail}。
 * rule 取值：schema | section | reference | runner
 *
 * @param {object} contract
 * @returns {{ok: boolean, errors: Array<{rule: string, detail: string}>}}
 */
function validateContract(contract) {
  const errors = [];
  const err = (rule, detail) => errors.push({ rule, detail });

  if (!contract || typeof contract !== 'object' || Array.isArray(contract)) {
    return { ok: false, errors: [{ rule: 'schema', detail: 'contract must be a plain object' }] };
  }
  if (contract.version !== CONTRACT_VERSION) {
    err('schema', `version must be "${CONTRACT_VERSION}", got ${JSON.stringify(contract.version)}`);
  }
  if (!contract.goal_id || typeof contract.goal_id !== 'string') {
    err('schema', 'missing field: goal_id');
  }
  for (const k of Object.keys(contract)) {
    if (!['version', 'goal_id', 'created_by', 'created_at', ...SECTIONS].includes(k)) {
      err('schema', `unknown top-level field: ${k}`);
    }
  }
  // created_by 两种形态都合法：单写者写字符串；三个产出节点各写一段时是 {section: producer} 映射
  // （§一.1「三个产出节点各写自己那一段」——一个字符串存不下三个作者）。审计字段，不参与判定。
  if (contract.created_by !== undefined && contract.created_by !== null) {
    const cb = contract.created_by;
    const okShape =
      typeof cb === 'string' ||
      (typeof cb === 'object' && !Array.isArray(cb) && Object.values(cb).every((v) => typeof v === 'string'));
    if (!okShape) err('schema', 'created_by must be a string or a {section: producer} object');
  }

  // 至少一个 section 非 null。空契约会「平凡通过」——正是 §四.D.3 要求 planner-validator
  // 盯的那类漏洞，在格式层面直接堵掉。
  const present = SECTIONS.filter((s) => contract[s] !== undefined && contract[s] !== null);
  if (present.length === 0) {
    err('section', 'at least one of benchmark / test_data / test_script must be non-null');
  }

  for (const s of present) {
    const sec = contract[s];
    if (typeof sec !== 'object' || Array.isArray(sec)) {
      err('section', `${s} must be an object or null`);
      continue;
    }
    // 未解引用的 $ref 只校验形态，内容等 readContract 解开后再校验
    if (sec.$ref !== undefined) {
      if (typeof sec.$ref !== 'string' || !sec.$ref) err('reference', `${s}.$ref must be a non-empty string`);
      else if (path.isAbsolute(sec.$ref) || sec.$ref.split('/').includes('..')) {
        err('reference', `${s}.$ref must be a relative path inside artifacts/: ${sec.$ref}`);
      }
      continue;
    }
    if (s === 'benchmark') validateBenchmark(sec, err);
    if (s === 'test_data') validateTestData(sec, err);
    if (s === 'test_script') validateTestScript(sec, err);
  }

  return { ok: errors.length === 0, errors };
}

function validateBenchmark(sec, err) {
  if (!Array.isArray(sec.dimensions) || sec.dimensions.length === 0) {
    err('section', 'benchmark.dimensions must be a non-empty array');
    return;
  }
  let sum = 0;
  const seen = new Set();
  for (const d of sec.dimensions) {
    if (!d || typeof d.name !== 'string' || !d.name) { err('section', 'benchmark.dimensions[].name required'); continue; }
    if (seen.has(d.name)) err('section', `benchmark duplicate dimension: ${d.name}`);
    seen.add(d.name);
    if (typeof d.weight !== 'number' || d.weight <= 0) err('section', `benchmark.${d.name}.weight must be > 0`);
    else sum += d.weight;
    if (d.pass_threshold !== undefined && typeof d.pass_threshold !== 'number') {
      err('section', `benchmark.${d.name}.pass_threshold must be a number`);
    }
  }
  if (seen.size > 0 && Math.abs(sum - 1) > 1e-6) {
    err('section', `benchmark dimension weights must sum to 1, got ${sum}`);
  }
  if (sec.overall_pass_threshold !== undefined && typeof sec.overall_pass_threshold !== 'number') {
    err('section', 'benchmark.overall_pass_threshold must be a number');
  }
}

function validateTestData(sec, err) {
  if (!Array.isArray(sec.cases) || sec.cases.length === 0) {
    err('section', 'test_data.cases must be a non-empty array');
    return;
  }
  if (sec.match !== undefined && !MATCH_MODES.includes(sec.match)) {
    err('section', `test_data.match must be one of ${MATCH_MODES.join('|')}`);
  }
  const seen = new Set();
  for (const c of sec.cases) {
    if (!c || typeof c.id !== 'string' || !c.id) { err('section', 'test_data.cases[].id required'); continue; }
    if (seen.has(c.id)) err('section', `test_data duplicate case id: ${c.id}`);
    seen.add(c.id);
    if (!('expect' in c)) err('section', `test_data.cases[${c.id}].expect required`);
  }
}

function validateTestScript(sec, err) {
  if (!RUNNERS.includes(sec.runner)) {
    err('runner', `test_script.runner must be one of ${RUNNERS.join('|')}, got ${JSON.stringify(sec.runner)}`);
  }
  if (typeof sec.path !== 'string' || !sec.path) {
    err('section', 'test_script.path required');
  } else if (path.isAbsolute(sec.path) || sec.path.split('/').includes('..')) {
    err('reference', `test_script.path must be a relative path inside artifacts/: ${sec.path}`);
  }
  if (sec.args !== undefined && !Array.isArray(sec.args)) err('section', 'test_script.args must be an array');
  if (sec.timeout_ms !== undefined && (typeof sec.timeout_ms !== 'number' || sec.timeout_ms <= 0)) {
    err('section', 'test_script.timeout_ms must be a positive number');
  }
}

/**
 * 解开 section 上的 $ref（相对 artifacts/，不许逃出去）。
 *
 * contract_hash 一律按**解引用后**的契约计算——否则「内联写」和「$ref 外置写」
 * 同一份验收标准会算出两个 hash，判决就追溯不回契约版本了。读和写都走这里。
 */
function resolveRefs(contract, artifactsDir) {
  const out = { ...contract };
  for (const s of SECTIONS) {
    const sec = out[s];
    if (!sec || typeof sec !== 'object' || sec.$ref === undefined) continue;
    const abs = resolveInsideArtifacts(artifactsDir, sec.$ref);
    if (!abs) throw new ContractValidationError([{ rule: 'reference', detail: `${s}.$ref escapes artifacts/: ${sec.$ref}` }]);
    const sub = readJsonFile(abs);
    if (!sub.ok) throw new ContractValidationError([{ rule: 'reference', detail: `${s}.$ref unreadable: ${sec.$ref} (${sub.error})` }]);
    out[s] = sub.json;
  }
  return out;
}

/**
 * 读契约并解开 $ref。
 * @returns {{contract: object, path: string}|null} 没有契约文件返回 null
 */
function readContract(runDir) {
  const p = resolveContractPath(runDir);
  if (!fs.existsSync(p)) return null;
  const r = readJsonFile(p);
  if (!r.ok) throw new ContractValidationError([{ rule: 'schema', detail: `unreadable ${CONTRACT_FILE}: ${r.error}` }]);
  return { contract: resolveRefs(r.json, resolveArtifactsDir(runDir)), path: p };
}

/**
 * 写契约（写前必校验，不通过抛 ContractValidationError；CONVENTIONS §12 同款约定）。
 * 文件按原样落盘（保留 $ref），返回的 contract_hash 按解引用后的形态算。
 * 因此 $ref 指向的文件必须先写好，否则这里就报错——比等到执行时才炸要早。
 */
function writeContract(runDir, contract) {
  const v = validateContract(contract);
  if (!v.ok) throw new ContractValidationError(v.errors);
  const resolved = resolveRefs(contract, resolveArtifactsDir(runDir));
  const rv = validateContract(resolved); // 解引用后的内容同样要合法
  if (!rv.ok) throw new ContractValidationError(rv.errors);
  const p = resolveContractPath(runDir);
  writeJsonAtomic(p, contract);
  return { path: p, contract_hash: hashContract(resolved) };
}

/**
 * 写**一段** section，合并进已有的 acceptance.json（产出侧唯一该用的入口）。
 *
 * §一.1 的不变量是「三个产出节点（goal-benchmark / goal-test-data / goal-test-script）
 * 各写自己那一段，互不覆盖」，但 `writeContract` 是整份替换——三个节点先后调它，
 * 后写的会把先写的那两段抹掉，契约到了 org-validator 手上只剩最后一段。
 * 这就是「产生 → 落库」之间被静默吞掉的那一环，所以产出侧必须走这个函数。
 *
 * @param {string} runDir
 * @param {object} args {section, content, goalId, createdBy}
 *        content 可以是内联对象，也可以是 {"$ref": "goal-benchmark.json"}
 * @returns {{path: string, contract_hash: string, sections: string[]}}
 */
function writeContractSection(runDir, { section, content, goalId, createdBy } = {}) {
  if (!SECTIONS.includes(section)) {
    throw new ContractValidationError([
      { rule: 'schema', detail: `section must be one of ${SECTIONS.join('|')}, got ${JSON.stringify(section)}` },
    ]);
  }
  if (!content || typeof content !== 'object' || Array.isArray(content)) {
    throw new ContractValidationError([{ rule: 'section', detail: `${section} content must be a plain object` }]);
  }

  // 读**原文**（不解 $ref）：合并要保留别人写的 $ref 形态，解开再写回等于把外置内容抄进契约
  const p = resolveContractPath(runDir);
  let base = {};
  if (fs.existsSync(p)) {
    const r = readJsonFile(p);
    if (!r.ok) throw new ContractValidationError([{ rule: 'schema', detail: `unreadable ${CONTRACT_FILE}: ${r.error}` }]);
    base = r.json && typeof r.json === 'object' && !Array.isArray(r.json) ? r.json : {};
  }

  const nextGoal = goalId || base.goal_id;
  if (base.goal_id && goalId && base.goal_id !== goalId) {
    // 一个 root run 一份契约、一个 goal（§一.1）。串了 goal 说明调用方拿错了 runDir，
    // 静默覆盖会让判决追溯到一个根本不是它的 goal_id 上。
    throw new ContractValidationError([
      { rule: 'schema', detail: `goal_id mismatch: existing ${base.goal_id}, incoming ${goalId}` },
    ]);
  }

  const merged = {
    ...base,
    version: CONTRACT_VERSION,
    goal_id: nextGoal,
    created_at: base.created_at || now(), // 保留首写时间：contract_hash 含它，刷新会让 hash 无谓变动
    [section]: content,
  };
  if (createdBy) {
    const prev = typeof base.created_by === 'string' ? {} : (base.created_by || {});
    merged.created_by = { ...prev, [section]: createdBy };
  }

  const out = writeContract(runDir, merged);
  return { ...out, sections: SECTIONS.filter((s) => merged[s] != null) };
}

// ────────────────────────────────────────────────────────────────
// 执行器
// ────────────────────────────────────────────────────────────────

/**
 * 短路失败（还没跑到逐维度/逐 case 比对就已经判定不通过）的 failures 条目。
 *
 * 形状与正常比对产出的 failure 完全一致（case / expected / got / hint），
 * 这样 `executeContract` 的 `failures: active.flatMap(...)` 和 CLI 的 `fmtFailures`
 * 都不用为短路分支分叉——**下游只认一种形状**。
 * `hint` 里带上 `reason`，让扁平汇总单看一行也能归因。
 */
function shortCircuitFailure(caseName, reason, detail, expected) {
  return {
    case: caseName,
    expected,
    got: detail === undefined || detail === null ? null : String(detail),
    hint: `${reason}: ${SHORT_CIRCUIT_HINTS[reason] || '执行前置条件不满足，未进入逐项比对'}`,
  };
}

const SHORT_CIRCUIT_HINTS = {
  target_not_json: '被测产物读不出 JSON（不存在 / 不是合法 JSON），无法比对',
  script_missing: '契约指向的验收脚本不在 artifacts/ 下，未起进程',
  script_escapes_artifacts: '验收脚本路径逃出 artifacts/，拒绝执行',
};

/**
 * benchmark：被测产物提供各维度分数，执行器只做加权与阈值比较（纯声明式，不起 LLM）。
 * 被测产物形态：{"scores": {"<dimension>": 87, ...}} 或直接 {"<dimension>": 87, ...}
 */
function runBenchmark(sec, target) {
  // 短路分支也必须进 failures：verdict.failures 是下游（validator agent / validator-triager）
  // 唯一的扁平归因入口，只留 reason 不留 failures 的话「读回」这一环就空了（§十 四环缺陷）。
  if (!target.ok) {
    return {
      ok: false, reason: 'target_not_json', detail: target.error,
      score: null, dimensions: [],
      failures: [shortCircuitFailure('benchmark', 'target_not_json', target.error, '可解析为 JSON 的被测产物')],
    };
  }
  const scores = (target.json && typeof target.json.scores === 'object' && target.json.scores) || target.json || {};
  const dims = [];
  let weighted = 0;
  let ok = true;

  for (const d of sec.dimensions) {
    const score = scores[d.name];
    const has = typeof score === 'number';
    const threshold = typeof d.pass_threshold === 'number' ? d.pass_threshold : null;
    const passed = has && (threshold === null || score >= threshold);
    if (!passed) ok = false;
    if (has) weighted += score * d.weight;
    dims.push({
      name: d.name,
      weight: d.weight,
      score: has ? score : null,
      pass_threshold: threshold,
      ok: passed,
      reason: has ? undefined : 'missing score in target',
    });
  }

  const overall = typeof sec.overall_pass_threshold === 'number' ? sec.overall_pass_threshold : null;
  const overallOk = overall === null || weighted >= overall;
  if (!overallOk) ok = false;

  return {
    ok,
    score: Number(weighted.toFixed(4)),
    overall_pass_threshold: overall,
    overall_ok: overallOk,
    dimensions: dims,
    failures: dims.filter((d) => !d.ok).map((d) => ({
      case: d.name,
      expected: d.pass_threshold,
      got: d.score,
      hint: d.reason || 'dimension score below pass_threshold',
    })),
  };
}

/**
 * test_data：被测产物提供各 case 的实际输出，执行器逐 case 比对（纯声明式，不起 LLM）。
 * 被测产物形态：{"results": {"<caseId>": <output>}} 或 {"results": [{"id":..,"output":..}]}
 */
function runTestData(sec, target) {
  if (!target.ok) {
    return {
      ok: false, reason: 'target_not_json', detail: target.error,
      match: sec.match || 'deep-equal',
      total: sec.cases.length, passed: 0, failed: sec.cases.length,
      failures: [shortCircuitFailure('test_data', 'target_not_json', target.error, '可解析为 JSON 的被测产物')],
    };
  }
  const raw = (target.json && target.json.results !== undefined ? target.json.results : target.json) || {};
  const actual = Array.isArray(raw)
    ? Object.fromEntries(raw.map((r) => [r && r.id, r && (r.output !== undefined ? r.output : r.actual)]))
    : raw;

  const mode = sec.match || 'deep-equal';
  const failures = [];
  for (const c of sec.cases) {
    const got = actual[c.id];
    const missing = !(c.id in actual);
    const passed = !missing && (mode === 'subset' ? matchSubset(c.expect, got) : deepEqual(c.expect, got));
    if (!passed) {
      failures.push({
        case: c.id,
        expected: c.expect,
        got: missing ? null : got,
        hint: missing ? `no result for case ${c.id} in target` : `case ${c.id} mismatch (${mode})`,
      });
    }
  }
  return {
    ok: failures.length === 0,
    match: mode,
    total: sec.cases.length,
    passed: sec.cases.length - failures.length,
    failed: failures.length,
    failures,
  };
}

/** 起子进程跑脚本；超时先 SIGTERM 再 SIGKILL（沿用 integrator-tools.killAll 的做法） */
function spawnScript({ cmd, args, cwd, env, timeoutMs }) {
  return new Promise((resolve) => {
    const started = Date.now();
    let stdout = '';
    let stderr = '';
    let timedOut = false;
    const child = spawn(cmd, args, { cwd, env, stdio: ['ignore', 'pipe', 'pipe'] });

    const cap = (buf, chunk) => (buf.length >= OUTPUT_CAP ? buf : buf + String(chunk).slice(0, OUTPUT_CAP - buf.length));
    child.stdout.on('data', (c) => { stdout = cap(stdout, c); });
    child.stderr.on('data', (c) => { stderr = cap(stderr, c); });

    const killer = setTimeout(() => {
      timedOut = true;
      try { child.kill('SIGTERM'); } catch { /* already gone */ }
      setTimeout(() => { try { child.kill('SIGKILL'); } catch { /* already gone */ } }, 3000).unref();
    }, timeoutMs);
    killer.unref();

    child.on('error', (e) => {
      clearTimeout(killer);
      resolve({ exit_code: null, spawn_error: e.message, stdout, stderr, timed_out: false, duration_ms: Date.now() - started });
    });
    child.on('close', (code, signal) => {
      clearTimeout(killer);
      resolve({ exit_code: code, signal: signal || null, stdout, stderr, timed_out: timedOut, duration_ms: Date.now() - started });
    });
  });
}

/**
 * test_script：真实子进程执行（DevLoop Integrator 起真实进程跑 e2e 的同款做法）。
 * 被测产物路径通过 env ACCEPTANCE_TARGET 交给脚本，脚本退出码即判决。
 */
async function runTestScript(sec, { artifactsDir, runDir, targetPath }) {
  const scriptAbs = resolveInsideArtifacts(artifactsDir, sec.path);
  if (!scriptAbs) {
    return {
      ok: false, reason: 'script_escapes_artifacts', detail: sec.path,
      runner: sec.runner, script: sec.path, exit_code: null,
      failures: [shortCircuitFailure(sec.path, 'script_escapes_artifacts', sec.path, 'artifacts/ 之内的相对脚本路径')],
    };
  }
  if (!fs.existsSync(scriptAbs)) {
    return {
      ok: false, reason: 'script_missing', detail: scriptAbs,
      runner: sec.runner, script: sec.path, exit_code: null,
      failures: [shortCircuitFailure(sec.path, 'script_missing', scriptAbs, '契约指向的脚本文件存在')],
    };
  }

  const res = await spawnScript({
    cmd: process.execPath, // RUNNERS 白名单目前只有 'node'
    args: [scriptAbs, ...(sec.args || [])],
    cwd: sec.cwd ? resolveInsideArtifacts(artifactsDir, sec.cwd) || runDir : runDir,
    env: {
      ...process.env,
      ACCEPTANCE_TARGET: targetPath,
      ACCEPTANCE_RUN_DIR: runDir || '',
      ACCEPTANCE_ARTIFACTS_DIR: artifactsDir || '',
    },
    timeoutMs: sec.timeout_ms || DEFAULT_SCRIPT_TIMEOUT_MS,
  });

  const expectCode = (sec.pass_on && typeof sec.pass_on.exit_code === 'number') ? sec.pass_on.exit_code : 0;
  const ok = !res.timed_out && !res.spawn_error && res.exit_code === expectCode;
  return {
    ok,
    runner: sec.runner,
    script: sec.path,
    exit_code: res.exit_code,
    signal: res.signal,
    timed_out: res.timed_out,
    duration_ms: res.duration_ms,
    stdout_excerpt: res.stdout.slice(-4000),
    stderr_excerpt: res.stderr.slice(-4000),
    reason: res.timed_out ? 'script_timeout' : (res.spawn_error ? 'spawn_error' : undefined),
    failures: ok ? [] : [{
      case: sec.path,
      expected: `exit ${expectCode}`,
      got: res.timed_out ? 'timeout' : (res.spawn_error || `exit ${res.exit_code}`),
      hint: (res.stderr || res.stdout).slice(-800),
    }],
  };
}

/**
 * 基础设施抖动 → infra_flake（§四.D.2.2 的取值域）。
 * 这类失败在引擎侧原地重跑、不烧任何计数，所以必须与「真的没达标」区分开。
 */
const INFRA_REASONS = new Set(['script_timeout', 'spawn_error']);

/**
 * 执行一份契约，产出判决（纯执行，不落盘）。
 *
 * @param {object} contract   已解引用的契约（readContract 的产物 / 手写对象）
 * @param {string} targetPath 被测产物绝对路径（通常来自 graph_run_nodes.output_ref）
 * @param {object} [opts]     {runDir, runId, nodeRunId, attempt}
 * @returns {Promise<object>} verdict
 */
async function executeContract(contract, targetPath, opts = {}) {
  const started = Date.now();
  const runDir = opts.runDir || null;
  const artifactsDir = runDir ? resolveArtifactsDir(runDir) : null;

  // 相对路径先归一：下面既拿它做 startsWith 判断又拿它做 path.relative，
  // 两处用不同形态会算出「判断说在 run 目录里、relative 却算出一串 ../」的结果。
  const targetAbs = path.resolve(targetPath);

  const base = {
    verdict_id: genVerdictId(),
    goal_id: contract && contract.goal_id,
    contract_hash: null,
    run_id: opts.runId || null,
    node_run_id: opts.nodeRunId || null,
    // ⚠️ 不能写 `opts.attempt || null`：attempt 是 **0 基**（graph-engine 的 picked.attempt 从 0 起），
    // 第一次尝试的 0 会被 `||` 判成假值吞掉，判决四元组当场少一环、追溯不回是哪次尝试。
    attempt: Number.isInteger(opts.attempt) ? opts.attempt : null,
    // 落在 run 目录里的产物存相对路径（整个 run 目录可整体搬走 / 归档后仍可读），
    // 外部路径原样存绝对路径
    target: runDir && targetAbs.startsWith(path.resolve(runDir) + path.sep)
      ? path.relative(path.resolve(runDir), targetAbs)
      : targetAbs,
    created_at: now(),
  };

  const v = validateContract(contract);
  if (!v.ok) {
    // 契约本身不合法：不是产物的错，归因为「输入约束没写清」，退回契约产出者。
    return {
      ...base,
      ok: false,
      fail_reason: 'input_underspecified',
      sections: { benchmark: null, test_data: null, test_script: null },
      failures: v.errors.map((e) => ({ case: e.rule, expected: 'valid contract', got: e.detail, hint: 'fix acceptance.json' })),
      duration_ms: Date.now() - started,
    };
  }
  base.contract_hash = hashContract(contract);

  const needsJson = contract.benchmark || contract.test_data;
  const target = needsJson
    ? (fs.existsSync(targetAbs)
      ? readJsonFile(targetAbs)
      : { ok: false, error: `target not found: ${targetAbs}` })
    : { ok: true, json: null };

  const sections = { benchmark: null, test_data: null, test_script: null };
  if (contract.benchmark) sections.benchmark = runBenchmark(contract.benchmark, target);
  if (contract.test_data) sections.test_data = runTestData(contract.test_data, target);
  if (contract.test_script) {
    // 脚本拿到的 ACCEPTANCE_TARGET 必须是绝对路径：子进程的 cwd 是 runDir/artifacts 下的某处，
    // 传相对路径过去它解析出来的是另一个文件。
    sections.test_script = await runTestScript(contract.test_script, { artifactsDir, runDir, targetPath: targetAbs });
  }

  const active = SECTIONS.map((s) => sections[s]).filter(Boolean);
  const ok = active.every((s) => s.ok);
  const infra = active.some((s) => INFRA_REASONS.has(s.reason));

  return {
    ...base,
    ok,
    // 判决本身只回答「这次验收有没有达成」，**不**寻址任何节点：它是机器执行的合同校验，
    // 手上只有 target 与 sections，没有图的信息。跨层定点（target_node → fail_node_ref）
    // 由下游 validator-triager 依据这里的 failures 归因产出（R-17.1 定案；原 R-16
    // 「回退只退到 org node 粒度」的假设已作废，引擎现在支持退到子图里的具体 node）。
    fail_reason: ok ? null : (infra ? 'infra_flake' : 'acceptance_not_met'),
    sections,
    // 扁平汇总，供下游 validator-triager 归因直接消费（R-17.2 定名）
    failures: active.flatMap((s) => s.failures || []),
    duration_ms: Date.now() - started,
  };
}

// ────────────────────────────────────────────────────────────────
// 判决持久化
// ────────────────────────────────────────────────────────────────

/**
 * 落盘一份判决：完整体写 verdicts/{verdictId}.json，摘要 append 到 verdicts/index.ndjson。
 * index 永不改已写行（CONVENTIONS §12 的 ndjson 不变量）。
 */
function persistVerdict(runDir, verdict) {
  const dir = resolveVerdictsDir(runDir);
  const p = path.join(dir, `${verdict.verdict_id}.json`);
  writeJsonAtomic(p, verdict);

  const indexPath = path.join(dir, 'index.ndjson');
  const entry = {
    verdict_id: verdict.verdict_id,
    ok: verdict.ok,
    fail_reason: verdict.fail_reason || null,
    goal_id: verdict.goal_id || null,
    run_id: verdict.run_id || null,
    node_run_id: verdict.node_run_id || null,
    // 同 executeContract：attempt 0 基，`||` 会把第一次尝试吞成 null，
    // 于是 listVerdicts 读回来的索引里首轮判决永远查不到 attempt。
    attempt: Number.isInteger(verdict.attempt) ? verdict.attempt : null,
    contract_hash: verdict.contract_hash || null,
    target: verdict.target || null,
    created_at: verdict.created_at,
    file: path.relative(runDir, p),
  };
  fs.appendFileSync(indexPath, JSON.stringify(entry) + '\n');
  return { path: p, indexPath, entry };
}

/** 查判决索引（可按 run / node / 只看失败过滤），新的在后，limit 取末尾 N 条 */
function listVerdicts(runDir, filter = {}) {
  const indexPath = path.join(resolveVerdictsDir(runDir), 'index.ndjson');
  if (!fs.existsSync(indexPath)) return [];
  let rows = fs
    .readFileSync(indexPath, 'utf8')
    .split('\n')
    .filter(Boolean)
    .map((l) => { try { return JSON.parse(l); } catch { return null; } })
    .filter(Boolean);
  if (filter.runId) rows = rows.filter((r) => r.run_id === filter.runId);
  if (filter.nodeRunId) rows = rows.filter((r) => r.node_run_id === filter.nodeRunId);
  if (filter.onlyFailed) rows = rows.filter((r) => !r.ok);
  if (filter.limit) rows = rows.slice(-filter.limit);
  return rows;
}

/** 读单份完整判决 */
function readVerdict(runDir, verdictId) {
  const p = path.join(resolveVerdictsDir(runDir), `${verdictId}.json`);
  if (!fs.existsSync(p)) return null;
  const r = readJsonFile(p);
  return r.ok ? r.json : null;
}

/**
 * 一站式入口：读 runDir 的契约 → 执行 → 落盘判决。引擎侧的 validator 节点调这个。
 *
 * @param {object} args {runDir, targetPath, runId, nodeRunId, attempt}
 * @returns {Promise<{verdict: object, path: string}>}
 */
async function runAcceptance({ runDir, targetPath, runId, nodeRunId, attempt }) {
  const loaded = readContract(runDir);
  if (!loaded) {
    throw new ContractValidationError([
      { rule: 'reference', detail: `no ${CONTRACT_FILE} under ${resolveArtifactsDir(runDir)}` },
    ]);
  }
  const verdict = await executeContract(loaded.contract, targetPath, { runDir, runId, nodeRunId, attempt });
  const { path: p } = persistVerdict(runDir, verdict);
  return { verdict, path: p };
}

// ────────────────────────────────────────────────────────────────
// CLI（agent 侧的生产入口，见文件顶部「生产调用方」）
// ────────────────────────────────────────────────────────────────

const USAGE = `goal-contract — 验收契约的产出 / 校验 / 执行 / 查询（27-goal-contract-format.md）

  put      --run-dir <D> --section <benchmark|test_data|test_script>
           (--file <X.json> | --ref <artifacts 下的相对路径>)
           [--goal-id <G>] [--created-by <节点本体名>]
           产出侧：把**一段** section 合并进 artifacts/acceptance.json，不覆盖别人写的段。

  validate --run-dir <D>
           只校验契约本身（空契约 / 权重不足 1 / runner 越界 / 路径逃逸都在这里被挡）。

  run      --run-dir <D> --target <被测产物路径>
           [--run-id <R>] [--node-run-id <N>] [--attempt <n>] [--json]
           消费侧：读契约 → 执行 → 落盘判决。缺省的 run-dir / node-run-id / attempt 会从 target 路径推。

  list     --run-dir <D> [--node-run-id <N>] [--only-failed] [--limit <n>]
  show     --run-dir <D> --verdict-id <V>

退出码：0 = 通过 / 操作成功；1 = **判决为 fail**（工具本身是好的）；2 = 跑不出判决（参数错 / 没契约 / 契约非法）。
`;

/** 从被测产物路径向上找到含 artifacts/acceptance.json 的那一层（agent 少填一个参数少错一次） */
function deriveRunDir(targetPath) {
  let dir = path.dirname(path.resolve(targetPath));
  for (let i = 0; i < 12; i++) {
    if (fs.existsSync(path.join(dir, 'artifacts', CONTRACT_FILE))) return dir;
    const up = path.dirname(dir);
    if (up === dir) break;
    dir = up;
  }
  return null;
}

/** .../nodes/{nodeRunId}/attempt-{n}.json → {nodeRunId, attempt}；推不出来就留 null */
function deriveTargetIds(targetPath) {
  const abs = path.resolve(targetPath);
  const parts = abs.split(path.sep);
  const i = parts.lastIndexOf('nodes');
  const nodeRunId = i >= 0 && parts[i + 1] ? parts[i + 1] : null;
  const m = /^attempt-(\d+)\./.exec(path.basename(abs));
  return { nodeRunId, attempt: m ? Number(m[1]) : null };
}

function fmtFailures(failures) {
  const rows = failures || [];
  const out = rows.slice(0, 20).map(
    (f) => `  - ${f.case}: expected=${JSON.stringify(f.expected)} got=${JSON.stringify(f.got)}${f.hint ? ` | ${f.hint}` : ''}`,
  );
  if (rows.length > 20) out.push(`  … 另有 ${rows.length - 20} 条，完整清单见 verdict 文件`);
  // 判决为 fail 却一条 failure 都没有，本身就是要报出来的异常（旧版直接打个空的 `failures:` 标题，
  // agent 看到的是「fail 但没有任何可归因信息」，分不清「工具吞了」还是「真的没细节」）。
  if (rows.length === 0) out.push('  （无逐项明细——看上面的 sections 行与 fail_reason）');
  return out;
}

/**
 * 一段 section 的人读摘要。
 *
 * ⚠️ **短路分支没有 score / passed / exit_code**（`target_not_json` / `script_missing` /
 * `script_escapes_artifacts` 直接 return，压根没跑到比对），旧版无条件打
 * `score=${s.benchmark.score}`，于是照着 `tools.md` 退出码表办事的 agent 只看到
 * 「benchmark=FAIL、score=undefined、failures 空」——而落盘的 verdict JSON 里
 * reason 和 detail 明明白白都在。**归因信息被吞在「读回」这一环**（§十 的四环缺陷模式）。
 * 所以这里按字段是否真的存在分别打，且**只要有 reason 就一定把 reason + detail 打出来**。
 */
function fmtSection(name, sec) {
  if (!sec) return [];
  const out = [];
  if (name === 'benchmark' && typeof sec.score === 'number') {
    out.push(`  benchmark  score=${sec.score} / ${sec.overall_pass_threshold ?? '-'}`);
  }
  if (name === 'test_data' && typeof sec.total === 'number') {
    out.push(`  test_data  ${sec.passed}/${sec.total} passed  (match=${sec.match})`);
  }
  // duration_ms 只有真起过进程才有：短路分支打一行 `exit=null` 会被读成「跑了、返回 null」
  if (name === 'test_script' && sec.duration_ms !== undefined) {
    out.push(`  test_script exit=${sec.exit_code} timed_out=${sec.timed_out === true} (${sec.duration_ms}ms)`);
  }
  if (sec.reason) {
    out.push(`  ${name}  reason=${sec.reason}`
      + (sec.detail === undefined || sec.detail === null ? '' : `  detail=${sec.detail}`));
  }
  if (name === 'test_script' && sec.stderr_excerpt) out.push(`  stderr: ${sec.stderr_excerpt.slice(-500)}`);
  if (out.length === 0) out.push(`  ${name}  ${sec.ok ? 'ok' : 'FAIL'}（判决文件里也没有更多明细）`);
  return out;
}

async function cliMain(argv) {
  const cmd = argv[0];
  if (!cmd || cmd === 'help' || cmd === '--help' || cmd === '-h') {
    process.stdout.write(USAGE);
    return cmd ? 0 : 2;
  }
  const { values: a } = require('util').parseArgs({
    args: argv.slice(1),
    allowPositionals: false,
    options: {
      'run-dir': { type: 'string' },
      target: { type: 'string' },
      'goal-id': { type: 'string' },
      section: { type: 'string' },
      file: { type: 'string' },
      ref: { type: 'string' },
      'created-by': { type: 'string' },
      'run-id': { type: 'string' },
      'node-run-id': { type: 'string' },
      attempt: { type: 'string' },
      'verdict-id': { type: 'string' },
      'only-failed': { type: 'boolean' },
      limit: { type: 'string' },
      json: { type: 'boolean' },
    },
  });

  const needRunDir = () => {
    const d = a['run-dir'] || (a.target ? deriveRunDir(a.target) : null);
    if (!d) throw new Error('missing --run-dir (且无法从 --target 推出 run 目录)');
    return d;
  };

  if (cmd === 'put') {
    const runDir = needRunDir();
    if (!a.file && !a.ref) throw new Error('put 需要 --file 或 --ref 其一');
    const content = a.ref ? { $ref: a.ref } : JSON.parse(fs.readFileSync(a.file, 'utf8'));
    const r = writeContractSection(runDir, {
      section: a.section, content, goalId: a['goal-id'], createdBy: a['created-by'],
    });
    process.stdout.write(`OK  ${a.section} 已写入 ${r.path}\n`
      + `contract_hash: ${r.contract_hash}\nsections: ${r.sections.join(', ') || '(空)'}\n`);
    return 0;
  }

  if (cmd === 'validate') {
    const runDir = needRunDir();
    const loaded = readContract(runDir); // $ref 解不开会在这里抛，正是要挡的
    if (!loaded) throw new Error(`没有 ${CONTRACT_FILE}：${resolveArtifactsDir(runDir)}`);
    const v = validateContract(loaded.contract);
    const present = SECTIONS.filter((s) => loaded.contract[s] != null);
    if (v.ok) {
      process.stdout.write(`OK  契约合法：${loaded.path}\nsections: ${present.join(', ')}\n`
        + `contract_hash: ${hashContract(loaded.contract)}\n`);
      return 0;
    }
    process.stdout.write('INVALID  契约不合法（= fail_reason: input_underspecified，退回契约产出者）\n'
      + v.errors.map((e) => `  - [${e.rule}] ${e.detail}`).join('\n') + '\n');
    return 1;
  }

  if (cmd === 'run') {
    if (!a.target) throw new Error('run 需要 --target <被测产物路径>');
    const runDir = needRunDir();
    const derived = deriveTargetIds(a.target);
    const attempt = a.attempt !== undefined ? Number(a.attempt) : derived.attempt;
    const { verdict, path: vp } = await runAcceptance({
      runDir,
      targetPath: a.target,
      runId: a['run-id'] || null,
      nodeRunId: a['node-run-id'] || derived.nodeRunId,
      attempt: Number.isInteger(attempt) ? attempt : null,
    });
    if (a.json) {
      process.stdout.write(JSON.stringify(verdict, null, 2) + '\n');
    } else {
      const s = verdict.sections;
      const lines = [
        `VERDICT ${verdict.ok ? 'pass' : 'fail'}  fail_reason=${verdict.fail_reason || 'null'}`
          + `  goal=${verdict.goal_id || '?'}  attempt=${verdict.attempt === null ? '?' : verdict.attempt}`,
        `verdict_id: ${verdict.verdict_id}   contract_hash: ${verdict.contract_hash || '?'}`,
        `verdict_file: ${vp}`,
        `target: ${verdict.target}`,
        `sections: ${SECTIONS.map((k) => (s[k] ? `${k}=${s[k].ok ? 'ok' : 'FAIL'}` : `${k}=-`)).join('  ')}`,
      ];
      for (const k of SECTIONS) lines.push(...fmtSection(k, s[k]));
      if (!verdict.ok) lines.push('failures:', ...fmtFailures(verdict.failures));
      process.stdout.write(lines.join('\n') + '\n');
    }
    return verdict.ok ? 0 : 1;
  }

  if (cmd === 'list') {
    const runDir = needRunDir();
    const rows = listVerdicts(runDir, {
      nodeRunId: a['node-run-id'], runId: a['run-id'],
      onlyFailed: !!a['only-failed'], limit: a.limit ? Number(a.limit) : undefined,
    });
    process.stdout.write(rows.length
      ? rows.map((r) => `${r.created_at}  ${r.ok ? 'pass' : 'fail'}  ${r.verdict_id}`
        + `  node=${r.node_run_id || '-'} attempt=${r.attempt === null ? '-' : r.attempt}`
        + `  ${r.fail_reason || ''}`).join('\n') + '\n'
      : '(没有判决记录)\n');
    return 0;
  }

  if (cmd === 'show') {
    const runDir = needRunDir();
    const v = readVerdict(runDir, a['verdict-id']);
    if (!v) throw new Error(`没有这份判决：${a['verdict-id']}`);
    process.stdout.write(JSON.stringify(v, null, 2) + '\n');
    return v.ok ? 0 : 1;
  }

  throw new Error(`未知子命令：${cmd}\n\n${USAGE}`);
}

if (require.main === module) {
  cliMain(process.argv.slice(2))
    .then((code) => process.exit(code))
    .catch((e) => {
      // 契约非法 / 参数错 / 没契约 → 2：与「判决为 fail」的 1 分开，
      // 否则 agent 分不清「产物没达标」和「工具压根没跑起来」（后者是 input_underspecified）。
      const detail = e instanceof ContractValidationError
        ? e.errors.map((x) => `  - [${x.rule}] ${x.detail}`).join('\n')
        : '';
      process.stderr.write(`ERROR  ${e.message}\n${detail}${detail ? '\n' : ''}`);
      process.exit(2);
    });
}

module.exports = {
  // 常量
  CONTRACT_VERSION,
  CONTRACT_FILE,
  SECTIONS,
  RUNNERS,
  ContractValidationError,
  // 路径
  resolveRunDir,
  resolveArtifactsDir,
  resolveContractPath,
  resolveVerdictsDir,
  // 契约
  validateContract,
  readContract,
  writeContract,
  writeContractSection,
  resolveRefs,
  contractHash: hashContract, // 入参必须是**已解引用**的契约

  // 执行
  executeContract,
  runAcceptance,
  // 判决
  persistVerdict,
  listVerdicts,
  readVerdict,
  // CLI（测试直接跑它，保证模板里印的那几行命令是真能跑的）
  cliMain,
  deriveRunDir,
  deriveTargetIds,
};