modelfunction-cli 0.5.4 → 0.5.5

package/README.md

@@ -5,7 +5,7 @@
 ## 安装
 
 ```bash
-npx -y modelfunction-cli@0.5.3 --version
+npx -y modelfunction-cli@0.5.5 --version
 # 或
 npm i -D modelfunction-cli
 npx mf --version
@@ -76,7 +76,7 @@ INTERNAL_LLM_KEY=... mf eval --suite mf.suite.json \
 | 命令 | 作用 |
 | --- | --- |
 | `mf scan` | 扫描工具定义（function.json / OpenAPI / MCP 配置 / Python `@tool` / JS·TS 声明与 TypeBox·zod schema）；`--probe` 连 stdio 或 streamable-HTTP MCP server（运行时结果覆盖静态推断；headers/query 不写入报告）；`--from` 导入运行时工具数组；草稿按来源分组（`profiles`）并列出未覆盖的 skill |
-| `mf eval` | 跑评分卡；`--repeat` 多次取多数并报不稳定用例；`--fail-under` 阈值；`--baseline` + `--fail-on-regression` 回归门禁；`--profile` 按挂载组分别评分；`--max-cost` 按累计成本停机（退出 2）；`--progress` 逐用例进度（stderr） |
+| `mf eval` | 跑评分卡；`--repeat` 多次取多数并报不稳定用例；`--fail-under` 阈值；`--baseline` + `--fail-on-regression` 回归门禁；`--profile` 按挂载组分别评分；`--max-cost` 按累计成本停机（退出 2）；`--retry` 429/5xx 重试次数；`--min-interval` 请求最小间隔（压 RPM）；`--progress` 逐用例进度（stderr） |
 | `mf case add` | 写入用例（默认 `source=agent`、`reviewed=false`） |
 | `mf coverage` | 每工具用例数、参数覆盖、负例比例、疑似混淆对；有 `profiles` 时按组统计 |
 | `mf audit` | 一次静态审计：lint + 混淆族边界 + 用例覆盖 + 挂载检查，输出可转发的 Markdown/JSON；不调用模型 |
@@ -126,21 +126,19 @@ mf eval --suite mf.suite.json --models internal-ds \
   "mcpServers": {
     "modelfunction": {
       "command": "npx",
-      "args": ["-y", "modelfunction-cli@0.5.3", "mcp"]
+      "args": ["-y", "modelfunction-cli@0.5.5", "mcp"]
     }
   }
 }
 ```
 
-暴露：`scan_tools`、`list_suite`、`coverage`、`audit`、`lint`、`check`、`cluster`、`compare`、`add_cases`、`run_eval`（`run_eval` 支持 `profile`、`repeat`、`maxCost`、`progress`）。`lint`/`check`/`cluster`/`compare` 都是静态操作，不调用模型；`check` 就是安装门禁，Agent 可直接传入 manifest 与 installed。
+暴露：`scan_tools`、`list_suite`、`coverage`、`audit`、`lint`、`check`、`cluster`、`compare`、`add_cases`、`run_eval`（`run_eval` 支持 `profile`、`repeat`、`maxCost`、`retry`、`minInterval`、`progress`）。`lint`/`check`/`cluster`/`compare` 都是静态操作，不调用模型；`check` 就是安装门禁，Agent 可直接传入 manifest 与 installed。
 
 ## 契约与版本
 
 对外 JSON 字段、`check()` 返回结构、MCP 参数与退出码见 [SCHEMA.md](SCHEMA.md)，自 0.4.3 起冻结为 v1：v1 内只做加法，重命名或删除字段需 major 并提前两个 minor 或 90 天弃用。`check()` 的类型声明随包发布。
 
-`0.5.3` 已发布到 npm：含凭证与本地反馈、usage `ci` 标记、eval `--progress`/`--max-cost`、`mf audit`、`scan --probe` streamable HTTP、MCP 静态工具 `lint`/`check`/`cluster`/`compare`、`checkMany` 一包批量检查、门禁严重度收敛、静态规则性能（真实 100 能力清单 `checkMany` 471ms→40ms）、`installed` 形状校验与 `profiles` 平台形态兼容。示例统一使用 `@0.5.3`。
-
-`0.5.4`（待发布）：用例新增 `excludeFromBase` / `excludeFromProfiles`（布尔，排除计分但不省调用）；评分卡新增行级 `errorKind` 与 `errors` / `errorsByKind` / `scoredTotal` / `selectionAccuracyExcludingErrors`；挂载对比不再把超时/HTTP 错误记成 jammed，分组表单列 `errors`；`mf audit` 分组表增加 `neg%`。契约见 `SCHEMA.md`。
+`0.5.5`（待 npm 发布）：eval 新增 `--retry <n>`（429/5xx/网络错误指数退避，默认 2，共 3 次尝试；优先采用服务端 `Retry-After`）与 `--min-interval <ms>`（请求最小间隔，用于压住端点 RPM），评分卡新增 `retries` 计数；CI 模板 static 层补上 `cluster --json`，与 `action.yml` 对齐；新增 `templates/local/` 与 `docs/local-gate.md`（内网/非 GitHub 环境的 pre-push 静态门禁、变更门、定时回归）。`0.5.4` 含用例级 `excludeFromBase` / `excludeFromProfiles`、评分卡 `errorKind` / `errorsByKind` / `scoredTotal` / `selectionAccuracyExcludingErrors`、挂载对比错误不计 jammed、audit 分组 `neg%` 与分组 token 用量聚合。`0.5.3` 及更早的能力见下：凭证与本地反馈、usage `ci` 标记、eval `--progress`/`--max-cost`、`mf audit`、`scan --probe` streamable HTTP、MCP 静态工具 `lint`/`check`/`cluster`/`compare`、`checkMany` 一包批量检查、门禁严重度收敛、静态规则性能（真实 100 能力清单 `checkMany` 471ms→40ms）、`installed` 形状校验与 `profiles` 平台形态兼容。示例统一使用 `@0.5.5`。
 
 ## 数据边界
 
@@ -169,7 +167,7 @@ mf feedback --out out.json  # 导出脱敏反馈包（聚合值，不含原始
 
 ```bash
 # 1. 把 mf 挂成 MCP server
-claude mcp add modelfunction -- npx -y modelfunction-cli@0.5.3 mcp
+claude mcp add modelfunction -- npx -y modelfunction-cli@0.5.5 mcp
 ```
 
 ```text

package/SCHEMA.md

@@ -95,19 +95,21 @@ skills 只做静态检查，不参与模型评测。
   "models": [ { "model": "dsv4-flash", "total": 12, "reviewedTotal": 12, "excludedTotal": 0,
                "selectionAccuracy": 0.83, "selectionAccuracyExcludingErrors": 0.83,
                "argAccuracy": 0.9, "passRate": 0.75, "errors": 0, "errorsByKind": {},
-               "scoredTotal": 12, "p50LatencyMs": 900,
+               "scoredTotal": 12, "retries": 0, "p50LatencyMs": 900,
                "p95LatencyMs": 1800, "avgCostCny": 0.001, "totalCostCny": 0.012,
                "inputTokens": 1000, "cachedTokens": 0, "outputTokens": 200,
                "rows": [ { "id": "c1", "expectedTool": "...", "actualTool": "...",
                            "selectionCorrect": true, "args": { "total": 1, "matched": 1, "failures": [] },
                            "passed": true, "reviewed": true, "source": "manual",
-                           "latencyMs": 900, "usage": { "input": 80, "output": 20 }, "error": null } ] } ],
+                           "latencyMs": 900, "usage": { "input": 80, "output": 20 }, "error": null,
+                           "retries": 1 } ] } ],
   "profiles": { "groups": [ ... ] }
 }
 ```
 
 - 正式指标只统计 `reviewed` 且未被 `excludeFromBase` 排除的行；未复核与排除结果单独给出。
 - `errors` 是调用层失败（超时、HTTP、网络）的数量，按 `errorsByKind` 分类；它们仍计入 `selectionAccuracy`（保守口径），同时给出排除错误后的 `selectionAccuracyExcludingErrors` 与可评分行数 `scoredTotal`。
+- `retries` 是本轮因 429/5xx/网络错误而重试的请求次数（模型级汇总；行级只在大于 0 时出现）。重试不改变判分：最终成功的请求按正常结果计分，最终失败的仍计入 `errors`。
 - `--repeat` 时模型对象带 `stability`（翻转用例与多次结果分布）。
 - `--profile` 时每个组含 `profile`、`mounted`、`allInOne`、`delta`、`lifted`、`jammed`、`errors`；错误行不参与挂载对比（不会被记成 jammed），每个组的 `reviewed` 是可评分行数。
 - `fingerprint` 由工具与用例内容计算；`mf compare` 用它判断两份报告是否可比。

package/lib/adapters/openai.js

@@ -1,5 +1,6 @@
 const TIMEOUT_MS = 60_000;
-const MAX_ATTEMPTS = 3;
+const DEFAULT_RETRIES = 2;
+const MAX_RETRY_DELAY_MS = 30_000;
 
 export function createOpenAiAdapter(config) {
   const apiKey = process.env[config.apiKeyEnv];
@@ -7,13 +8,25 @@ export function createOpenAiAdapter(config) {
     throw new Error(`${config.id}: environment variable ${config.apiKeyEnv} is not set`);
   }
   const endpoint = `${config.baseUrl.replace(/\/+$/, "")}/chat/completions`;
+  const retries = clampInt(config.retries, 0, 10, DEFAULT_RETRIES);
+  const minIntervalMs = clampInt(config.minIntervalMs, 0, 60_000, 0);
+  let lastRequestAt = 0;
+
+  async function throttle() {
+    if (minIntervalMs <= 0) return;
+    const wait = lastRequestAt + minIntervalMs - Date.now();
+    if (wait > 0) await sleep(wait);
+    lastRequestAt = Date.now();
+  }
 
   return {
     label: config.model,
     async call({ messages, tools }) {
       const started = performance.now();
       let lastError = "request failed";
-      for (let attempt = 1; attempt <= MAX_ATTEMPTS; attempt += 1) {
+      for (let attempt = 1; attempt <= retries + 1; attempt += 1) {
+        const retried = attempt - 1;
+        await throttle();
         try {
           const response = await fetch(endpoint, {
             method: "POST",
@@ -33,13 +46,16 @@ export function createOpenAiAdapter(config) {
 
           if (response.status === 429 || response.status >= 500) {
             lastError = `HTTP ${response.status}: ${(await response.text()).slice(0, 300)}`;
-            await sleep(attempt * 1000);
+            if (attempt <= retries) {
+              await sleep(retryDelayMs(response.headers.get("retry-after"), retried));
+            }
             continue;
           }
           if (!response.ok) {
             const body = await response.text();
             return {
               ok: false,
+              retries: retried,
               error: `HTTP ${response.status}: ${body.slice(0, 300)}`,
               toolCalls: [],
               usage: { input: 0, output: 0 },
@@ -55,6 +71,7 @@ export function createOpenAiAdapter(config) {
           }));
           return {
             ok: true,
+            retries: retried,
             toolCalls,
             usage: {
               input: data.usage?.prompt_tokens ?? 0,
@@ -66,11 +83,12 @@ export function createOpenAiAdapter(config) {
           };
         } catch (error) {
           lastError = error?.name === "TimeoutError" ? `timeout after ${TIMEOUT_MS}ms` : String(error?.message ?? error);
-          if (attempt < MAX_ATTEMPTS) await sleep(attempt * 1000);
+          if (attempt <= retries) await sleep(retryDelayMs(null, retried));
         }
       }
       return {
         ok: false,
+        retries,
         error: lastError,
         toolCalls: [],
         usage: { input: 0, output: 0 },
@@ -80,6 +98,22 @@ export function createOpenAiAdapter(config) {
   };
 }
 
+function retryDelayMs(retryAfter, retried) {
+  const header = retryAfter == null || String(retryAfter).trim() === "" ? NaN : Number(retryAfter);
+  if (Number.isFinite(header) && header >= 0) {
+    return Math.min(header * 1000, MAX_RETRY_DELAY_MS);
+  }
+  return Math.min(1000 * 2 ** retried, MAX_RETRY_DELAY_MS);
+}
+
+function clampInt(value, min, max, fallback) {
+  const parsed = Number(value);
+  if (!Number.isFinite(parsed)) return fallback;
+  const rounded = Math.floor(parsed);
+  if (rounded < min || rounded > max) return fallback;
+  return rounded;
+}
+
 function parseArguments(raw) {
   if (raw === undefined || raw === null || raw === "") return {};
   if (typeof raw === "object") return raw;

package/lib/exec.js

@@ -1,4 +1,5 @@
 import { spawn } from "node:child_process";
+import { realpathSync } from "node:fs";
 import { mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { dirname, join, resolve } from "node:path";
@@ -13,18 +14,19 @@ const ENV_KEEP = ["PATH", "SystemRoot", "SYSTEMROOT", "TEMP", "TMP", "HOME", "US
 export async function executeTool(options) {
   const runtime = options.runtime ?? "node";
   const engine = runtime === "python" ? "python-audit" : "node-permission";
-  const entry = resolve(options.entry);
+  const entry = realpathAllow(options.entry);
   const nodeMajor = Number(process.versions.node.split(".")[0]);
   if (runtime === "node" && options.allowNet?.length && nodeMajor < 25) {
     throw new Error("--allow-net 需要 Node >= 25（Node 24 的权限模型不限制网络）");
   }
 
   let tempDir = null;
-  let inputPath = options.inputPath ? resolve(options.inputPath) : null;
+  let inputPath = options.inputPath ? realpathAllow(options.inputPath) : null;
   if (!inputPath) {
     tempDir = await mkdtemp(join(tmpdir(), "mf-exec-"));
     inputPath = join(tempDir, "input.json");
     await writeFile(inputPath, options.inputJson ?? "{}", "utf8");
+    inputPath = realpathAllow(inputPath);
   }
 
   try {
@@ -57,11 +59,11 @@ export async function executeTool(options) {
         dirname(inputPath),
         process.cwd(),
         ...(options.allowRead ?? [])
-      ].map((path) => resolve(path));
+      ].map((path) => realpathAllow(path));
       args = [permissionFlag()];
       for (const path of dedupe(readPaths)) args.push(`--allow-fs-read=${path}`);
       if (options.allowWrite?.length) {
-        for (const path of dedupe(options.allowWrite.map((item) => resolve(item)))) {
+        for (const path of dedupe(options.allowWrite.map((item) => realpathAllow(item)))) {
           args.push(`--allow-fs-write=${path}`);
         }
       }
@@ -264,3 +266,14 @@ function formatExec(entry, result) {
   if (result.permissionDenied) lines.push("  note:       权限模型拒绝了文件/网络访问，可用于推断沙箱需求");
   return lines.join("\n");
 }
+
+// Node's permission model matches real paths. On macOS /var is a symlink to
+// /private/var, so an unresolved tmpdir path in --allow-fs-read is rejected.
+function realpathAllow(path) {
+  const abs = resolve(path);
+  try {
+    return realpathSync.native ? realpathSync.native(abs) : realpathSync(abs);
+  } catch {
+    return abs;
+  }
+}

package/lib/mcp-server.js

@@ -248,10 +248,12 @@ export async function startMcpServer() {
           .describe("每条用例重复次数，取多数结果；建议正式基线用 3"),
         failUnder: z.number().optional().describe("选择准确率阈值，低于则退出码为 2"),
         maxCost: z.number().positive().optional().describe("累计估算成本上限（CNY），超限退出 2 且不返回部分评分卡"),
+        retry: z.number().int().min(0).max(10).optional().describe("单次请求遇到 429/5xx/网络错误的重试次数（默认 2，共 3 次尝试）"),
+        minInterval: z.number().int().min(0).max(60000).optional().describe("两次模型请求的最小间隔（毫秒），用于压住端点 RPM"),
         progress: z.boolean().optional().describe("把逐用例进度写到 stderr（stdout 仍为纯 JSON）")
       }
     },
-    async ({ suite, models, cases, profile, repeat, failUnder, maxCost, progress }) => {
+    async ({ suite, models, cases, profile, repeat, failUnder, maxCost, retry, minInterval, progress }) => {
       const args = ["eval", "--json", "--suite", suitePath(suite)];
       if (models) args.push("--models", models);
       if (cases) args.push("--case", cases);
@@ -259,6 +261,8 @@ export async function startMcpServer() {
       if (repeat !== undefined) args.push("--repeat", String(repeat));
       if (failUnder !== undefined) args.push("--fail-under", String(failUnder));
       if (maxCost !== undefined) args.push("--max-cost", String(maxCost));
+      if (retry !== undefined) args.push("--retry", String(retry));
+      if (minInterval !== undefined) args.push("--min-interval", String(minInterval));
       if (progress) args.push("--progress");
 
       const result = await runCli(args);

package/lib/score.js

@@ -112,6 +112,7 @@ export function scoreModel({ model, cases, results, pricing, scope = "base" }) {
       source: testCase.source ?? "manual",
       latencyMs: result.latencyMs ?? 0,
       usage: result.usage ?? { input: 0, output: 0 },
+      ...(result.retries > 0 ? { retries: result.retries } : {}),
       error,
       ...(error ? { errorKind: classifyError(error) } : {})
     };
@@ -162,6 +163,7 @@ export function scoreModel({ model, cases, results, pricing, scope = "base" }) {
     errors,
     errorsByKind,
     scoredTotal,
+    retries: rows.reduce((sum, row) => sum + (row.retries ?? 0), 0),
     p50LatencyMs: percentile(latencies, 0.5),
     p95LatencyMs: percentile(latencies, 0.95),
     avgCostCny: total ? costCny / total : 0,

package/mf.js

@@ -97,6 +97,8 @@ Eval options:
   --profile <names>     score each mounted group separately ("all" or names from suite.profiles)
   --repeat <n>          run each case n times and score the majority; reports unstable cases (default 1)
   --max-cost <cny>      stop after the estimated spend crosses this limit (exit 2; no partial scorecard)
+  --retry <n>           retries per request on HTTP 429/5xx and network errors (default 2; 3 attempts total)
+  --min-interval <ms>   minimum spacing between model requests; use it to stay under an endpoint RPM
   --progress            print per-case progress to stderr (stdout stays clean for --json)
   --show-cases          print per-case results
   --json                print JSON instead of the table
@@ -136,7 +138,10 @@ export async function runEval(options, io = console) {
   }
   const suite = await loadSuite(options.suite);
   const cases = selectCases(suite.cases, options.cases);
-  const { registry, adapters } = await resolveModels(options.models, options.modelsFile, io);
+  const { registry, adapters } = await resolveModels(options.models, options.modelsFile, io, {
+    retries: options.retries,
+    minIntervalMs: options.minIntervalMs
+  });
   const tools = toModelTools(suite.tools);
 
   const budget = createBudget(options.maxCost ?? null);
@@ -393,7 +398,10 @@ export async function runAblate(options, io = console) {
     : selectedCases;
   const reducedCases = baseCases.filter((testCase) => testCase.expectedTool !== ablated.name);
   const droppedCases = baseCases.length - reducedCases.length;
-  const { registry, adapters } = await resolveModels(options.models, options.modelsFile, io);
+  const { registry, adapters } = await resolveModels(options.models, options.modelsFile, io, {
+    retries: options.retries,
+    minIntervalMs: options.minIntervalMs
+  });
 
   const fullTools = toModelTools(baseTools);
   const reducedTools = toModelTools(baseTools.filter((tool) => tool.name !== ablated.name));
@@ -778,7 +786,7 @@ function selectCases(cases, ids) {
   return cases.filter((testCase) => wanted.has(testCase.id));
 }
 
-async function resolveModels(ids, modelsFile, io) {
+async function resolveModels(ids, modelsFile, io, network = {}) {
   const registry = JSON.parse(await readFile(resolve(modelsFile), "utf8"));
   // Env wins; the credentials file only fills what is missing.
   await applyModelCredentials(registry);
@@ -805,7 +813,15 @@ async function resolveModels(ids, modelsFile, io) {
       continue;
     }
     registerSecrets([apiKey]);
-    adapters.push({ id, ...createOpenAiAdapter({ id, ...config }) });
+    adapters.push({
+      id,
+      ...createOpenAiAdapter({
+        id,
+        ...config,
+        retries: network.retries,
+        minIntervalMs: network.minIntervalMs
+      })
+    });
   }
 
   if (adapters.length === 0) {
@@ -1032,6 +1048,8 @@ function parseArgs(argv) {
     profiles: null,
     repeat: 1,
     maxCost: null,
+    retries: 2,
+    minIntervalMs: 0,
     progress: false,
     showCases: false,
     json: false,
@@ -1097,6 +1115,18 @@ function parseArgs(argv) {
         options.maxCost = raw;
         break;
       }
+      case "--retry": {
+        const raw = Number(value());
+        if (!Number.isInteger(raw) || raw < 0 || raw > 10) throw new Error("--retry expects an integer between 0 and 10");
+        options.retries = raw;
+        break;
+      }
+      case "--min-interval": {
+        const raw = Number(value());
+        if (!Number.isInteger(raw) || raw < 0 || raw > 60_000) throw new Error("--min-interval expects an integer between 0 and 60000 (ms)");
+        options.minIntervalMs = raw;
+        break;
+      }
       case "--progress": options.progress = true; break;
       case "--show-cases": options.showCases = true; break;
       case "--json": options.json = true; break;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "modelfunction-cli",
-  "version": "0.5.4",
+  "version": "0.5.5",
   "type": "module",
   "license": "MIT",
   "description": "mf — function-calling accuracy scorecard for AI agents",