modelfunction-cli 0.0.0-stage → 0.1.1

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Shanghai Yunzerui Technology Co., Ltd.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -1,3 +1,119 @@
-# Temporary Holding Version
+# modelfunction-cli
 
-This version is a temporary placeholder for this package. An operational version to replace this has been submitted for review and is awaiting a staged release.
+`mf` — 测你的模型会不会选对 function、参数填得对不对，输出多模型评分卡。本地运行，不开机自启、不驻留、不上传代码。
+
+## 安装
+
+```bash
+npx -y modelfunction-cli@0.1.1 --version
+# 或
+npm i -D modelfunction-cli
+npx mf --version
+```
+
+## 五步上手
+
+```bash
+mf scan --root .                      # 找 function.json / OpenAPI / MCP 配置 / SKILL.md
+mf scan --root . --probe              # 连 stdio MCP server 拉 tools/list
+mf coverage --suite mf.suite.json     # 看每个工具还缺多少用例
+mf eval --suite mf.suite.json --models deepseek,qwen --show-cases
+mf compare before.json after.json     # 跨版本比较（内容指纹判定可比性）
+```
+
+被测模型用环境变量提供 Key：
+
+```bash
+DEEPSEEK_API_KEY=... mf eval --models deepseek
+DASHSCOPE_API_KEY=... mf eval --models qwen
+ZHIPU_API_KEY=... mf eval --models glm
+SILICONFLOW_API_KEY=... mf eval --models dsv4-flash
+OPENAI_API_KEY=... mf eval --models openai
+```
+
+### 自建 / 内网网关
+
+企业内网基本都是 OpenAI 兼容协议，写一个 models-file 即可，Key 只走环境变量、不落盘：
+
+```json
+{
+  "internal-ds": {
+    "type": "openai",
+    "baseUrl": "https://llm.example.internal/v1",
+    "model": "deepseek-v4-flash",
+    "apiKeyEnv": "INTERNAL_LLM_KEY"
+  }
+}
+```
+
+```bash
+INTERNAL_LLM_KEY=... mf eval --suite mf.suite.json \
+  --models-file my-models.json --models internal-ds
+```
+
+`pricing` 可省（省了就不输出成本）；要算成本请按自家价目填写，默认 registry 里的价格只是参考值。
+
+## 命令
+
+| 命令 | 作用 |
+| --- | --- |
+| `mf scan` | 扫描工具定义；`--probe` 连 MCP server；`--from` 导入运行时导出的工具数组 |
+| `mf eval` | 跑评分卡；`--fail-under` 阈值；`--baseline` + `--fail-on-regression` 回归门禁 |
+| `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 |
+| `mf lint` | 静态检查描述与参数 schema |
+| `mf compare` | 跨版本比较、干扰归因、混淆矩阵 |
+| `mf ablate` | 留一法：去掉某个工具再跑一遍，量化它是否干扰其他工具 |
+| `mf exec` | 单工具执行 + 指标（Node / Python），支持 `--allow-read/write/net`、超时、sideEffects 准入 |
+| `mf eval --execute` | 一次跑完选择 + 参数 + 运行断言（`toolInput` / `expect`） |
+| `mf sandbox check` | 探测本机可用隔离引擎与降级建议 |
+| `mf mcp` | stdio MCP server，供 Claude Code / Cursor / Kimi Code 调用 |
+
+## MCP 配置
+
+```json
+{
+  "mcpServers": {
+    "modelfunction": {
+      "command": "npx",
+      "args": ["-y", "modelfunction-cli@0.1.1", "mcp"]
+    }
+  }
+}
+```
+
+暴露：`scan_tools`、`list_suite`、`coverage`、`add_cases`、`run_eval`。
+
+## 数据边界
+
+- 本地运行，无遥测、无服务端地址
+- 唯一出站：你配置的模型端点，以及你自己的 MCP server（`--probe` 时）
+- 评测会把用例输入与工具描述发给被测模型供应商——敏感数据请选境内供应商或本地模型
+
+## 用例从哪来
+
+由你或你的编码 Agent 撰写，规则模板见包内 `skills/mf-cases/SKILL.md`。未复核用例（`reviewed:false`）不计入正式分数。
+
+工具一多就别手工誊抄，直接让编码 Agent 驱动 `mf`：
+
+```bash
+# 1. 把 mf 挂成 MCP server
+claude mcp add modelfunction -- npx -y modelfunction-cli@0.1.1 mcp
+```
+
+```text
+# 2. 让 Agent 扫仓库并写出套件草稿
+用 modelfunction 的 scan_tools 扫 .，probe=true，writeSuite=mf.suite.json
+
+# 3. 让 Agent 补用例
+按 skills/mf-cases/SKILL.md 的规则，为每个工具补 3 条正例和 1 条负例，用 add_cases
+写入 mf.suite.json，然后跑 coverage 看还缺哪些参数和负例
+```
+
+```bash
+# 4. 人工复核：把认可的用例 reviewed 改为 true，再跑正式评分
+mf coverage --suite mf.suite.json
+mf eval --suite mf.suite.json --models internal-ds
+```
+
+License: MIT

package/fixtures/models.json

@@ -0,0 +1,39 @@
+{
+  "_note": "pricing 为人民币/百万 tokens 的参考值，运行前请按各家最新价目核对",
+  "mock": { "type": "mock" },
+  "deepseek": {
+    "type": "openai",
+    "baseUrl": "https://api.deepseek.com/v1",
+    "model": "deepseek-chat",
+    "apiKeyEnv": "DEEPSEEK_API_KEY",
+    "pricing": { "input": 2, "output": 8 }
+  },
+  "dsv4-flash": {
+    "type": "openai",
+    "baseUrl": "https://api.siliconflow.cn/v1",
+    "model": "deepseek-ai/DeepSeek-V4-Flash",
+    "apiKeyEnv": "SILICONFLOW_API_KEY",
+    "pricing": { "input": 1.0, "cachedInput": 0.02, "output": 2.0 }
+  },
+  "qwen": {
+    "type": "openai",
+    "baseUrl": "https://dashscope.aliyuncs.com/compatible-mode/v1",
+    "model": "qwen-plus",
+    "apiKeyEnv": "DASHSCOPE_API_KEY",
+    "pricing": { "input": 0.8, "output": 2 }
+  },
+  "glm": {
+    "type": "openai",
+    "baseUrl": "https://open.bigmodel.cn/api/paas/v4",
+    "model": "glm-4-air",
+    "apiKeyEnv": "ZHIPU_API_KEY",
+    "pricing": { "input": 0.5, "output": 0.5 }
+  },
+  "openai": {
+    "type": "openai",
+    "baseUrl": "https://api.openai.com/v1",
+    "model": "gpt-4o-mini",
+    "apiKeyEnv": "OPENAI_API_KEY",
+    "pricing": { "input": 1.1, "output": 4.3 }
+  }
+}

package/fixtures/suite.json

@@ -0,0 +1,150 @@
+{
+  "name": "recruiting-agent",
+  "description": "招聘 Agent 的 function-calling 评测集：简历评分、面试题、薪资区间、JD 改写。",
+  "tools": [
+    { "ref": "./tools/resume_score.json" },
+    {
+      "name": "interview_question",
+      "description": "根据岗位和候选人背景生成针对性的模拟面试题。",
+      "parameters": {
+        "type": "object",
+        "required": ["role"],
+        "properties": {
+          "role": { "type": "string", "description": "岗位名称，如后端工程师" },
+          "background": { "type": "string", "description": "候选人简历或背景摘要" },
+          "count": { "type": "number", "description": "题目数量，默认 5" }
+        }
+      }
+    },
+    {
+      "name": "salary_band",
+      "description": "根据城市、岗位和年限给出薪资区间参考。",
+      "parameters": {
+        "type": "object",
+        "required": ["city", "role"],
+        "properties": {
+          "city": { "type": "string", "description": "城市，如上海" },
+          "role": { "type": "string", "description": "岗位名称" },
+          "years": { "type": "number", "description": "工作年限" }
+        }
+      }
+    },
+    {
+      "name": "jd_rewrite",
+      "description": "把原始职位描述改写得更有吸引力，保留关键要求。",
+      "parameters": {
+        "type": "object",
+        "required": ["jd"],
+        "properties": {
+          "jd": { "type": "string", "description": "原始职位描述文本" },
+          "tone": { "type": "string", "enum": ["专业", "活泼", "简洁"], "description": "语气风格" }
+        }
+      }
+    }
+  ],
+  "cases": [
+    {
+      "id": "r01",
+      "input": "帮我给这个候选人打分，看看和岗位的匹配度。\n\n简历：张三，5 年 Python 后端经验，做过 RAG 检索增强和 LLM 应用，熟悉 MySQL、Redis、Docker，带过 3 人小组。\n\n职位描述：招聘 AI 应用后端工程师，3 年以上经验，要求熟悉 Python、MySQL、Redis、Docker、Kubernetes，有 LLM 与 RAG 项目经验优先。",
+      "expectedTool": "resume_score",
+      "expectedArgs": {
+        "resume": { "$contains": "张三" },
+        "jd": { "$contains": "Kubernetes" }
+      }
+    },
+    {
+      "id": "r02",
+      "input": "Do these match?\n\nResume: Bob, 6 years of Python and FastAPI experience, built RAG pipelines with Redis and Docker.\n\nJD: Backend engineer, 3+ years Python, Redis, Docker, Kubernetes, LLM experience preferred.",
+      "expectedTool": "resume_score",
+      "expectedArgs": {
+        "resume": { "$contains": "Bob" },
+        "jd": { "$contains": "Kubernetes" }
+      }
+    },
+    {
+      "id": "r03",
+      "input": "看看这个候选人和岗位的匹配度怎么样？",
+      "expectedTool": null,
+      "note": "没有简历或 JD 内容，应追问，而不是用编造的参数调用工具"
+    },
+    {
+      "id": "q01",
+      "input": "候选人简历：李四，3 年前端开发，做过 React 和 Vue 项目，主导过一个低代码平台。\n岗位：高级前端工程师\n帮我出 5 道面试题。",
+      "expectedTool": "interview_question",
+      "expectedArgs": {
+        "background": { "$contains": "李四" },
+        "role": { "$contains": "前端" },
+        "count": 5
+      }
+    },
+    {
+      "id": "q02",
+      "input": "给我出几道关于 Redis 和消息队列的面试题。",
+      "expectedTool": "interview_question",
+      "expectedArgs": {},
+      "note": "只测工具选择：没有岗位和背景信息"
+    },
+    {
+      "id": "s01",
+      "input": "城市：上海\n岗位：后端工程师\n年限：5\n这个岗位大概能给多少薪水？",
+      "expectedTool": "salary_band",
+      "expectedArgs": {
+        "city": { "$contains": "上海" },
+        "role": { "$contains": "后端" },
+        "years": 5
+      }
+    },
+    {
+      "id": "s02",
+      "input": "我们在杭州招一个算法工程师，3 年经验，预算大概多少？",
+      "expectedTool": "salary_band",
+      "expectedArgs": {
+        "city": { "$contains": "杭州" },
+        "role": { "$contains": "算法" },
+        "years": 3
+      }
+    },
+    {
+      "id": "j01",
+      "input": "岗位：AI 产品经理\n职位描述：负责 AI 产品规划与落地，要求 3 年 B 端产品经验，熟悉大模型能力边界。把它改写得更吸引人，语气活泼一点。",
+      "expectedTool": "jd_rewrite",
+      "expectedArgs": {
+        "jd": { "$contains": "AI 产品规划" },
+        "tone": "活泼"
+      }
+    },
+    {
+      "id": "j02",
+      "input": "帮我把这段职位描述润色得更简洁。\n职位描述：我们寻找一位熟悉 Kubernetes 和 Prometheus 的 SRE 工程师。",
+      "expectedTool": "jd_rewrite",
+      "expectedArgs": {
+        "jd": { "$contains": "Kubernetes" }
+      }
+    },
+    {
+      "id": "n01",
+      "input": "你好，你能做什么？",
+      "expectedTool": null
+    },
+    {
+      "id": "n02",
+      "input": "你们怎么收费？",
+      "expectedTool": null
+    },
+    {
+      "id": "n03",
+      "input": "帮我写一封邮件给候选人，说明面试安排。",
+      "expectedTool": null,
+      "note": "没有邮件工具，不应误调用面试题工具"
+    },
+    {
+      "id": "x01",
+      "input": "简历里有 5 年经验，这个岗位在上海能给多少薪水？",
+      "expectedTool": "salary_band",
+      "expectedArgs": {
+        "city": { "$contains": "上海" }
+      },
+      "note": "有“简历”字样，但真实意图是薪资"
+    }
+  ]
+}

package/fixtures/tools/resume_score.json

@@ -0,0 +1,37 @@
+{
+  "name": "resume_score",
+  "description": "简历与 JD 匹配评分",
+  "version": "0.1.0",
+  "input": {
+    "type": "object",
+    "required": ["resume", "jd"],
+    "properties": {
+      "resume": {
+        "type": "string",
+        "minLength": 20,
+        "description": "候选人简历文本"
+      },
+      "jd": {
+        "type": "string",
+        "minLength": 20,
+        "description": "职位描述文本"
+      }
+    }
+  },
+  "output": {
+    "type": "object",
+    "properties": {
+      "score": { "type": "number", "minimum": 0, "maximum": 100 },
+      "reasons": { "type": "array", "items": { "type": "string" } },
+      "matched": { "type": "array", "items": { "type": "string" } },
+      "missing": { "type": "array", "items": { "type": "string" } },
+      "years": { "type": "number" }
+    }
+  },
+  "sideEffects": "none",
+  "model": "mock-keyword-v1",
+  "price": { "amount": 0.02, "currency": "CNY", "unit": "call" },
+  "mcp": true,
+  "auth": { "type": "api_key" },
+  "rateLimit": { "requests": 60, "window": "1m" }
+}

package/lib/ablate.js

@@ -0,0 +1,120 @@
+export function analyzeAblation(model, full, reduced) {
+  const fullRows = new Map((full.rows ?? []).map((row) => [row.id, row]));
+  const reducedRows = new Map((reduced.rows ?? []).map((row) => [row.id, row]));
+  const comparedIds = [...reducedRows.keys()].filter((id) => fullRows.has(id));
+
+  const improved = [];
+  const worsened = [];
+  const perTool = new Map();
+
+  for (const id of comparedIds) {
+    const rowFull = fullRows.get(id);
+    const rowReduced = reducedRows.get(id);
+    if (rowFull.reviewed === false || rowReduced.reviewed === false) continue;
+
+    const label = rowFull.expectedTool ?? "(none)";
+    const bucket = perTool.get(label) ?? { tool: label, total: 0, fullCorrect: 0, reducedCorrect: 0 };
+    bucket.total += 1;
+    if (rowFull.selectionCorrect) bucket.fullCorrect += 1;
+    if (rowReduced.selectionCorrect) bucket.reducedCorrect += 1;
+    perTool.set(label, bucket);
+
+    if (rowFull.selectionCorrect && !rowReduced.selectionCorrect) {
+      worsened.push({
+        id,
+        expected: rowFull.expectedTool,
+        fullTool: rowFull.actualTool ?? null,
+        reducedTool: rowReduced.actualTool ?? null
+      });
+    }
+    if (!rowFull.selectionCorrect && rowReduced.selectionCorrect) {
+      improved.push({
+        id,
+        expected: rowFull.expectedTool,
+        fullTool: rowFull.actualTool ?? null,
+        reducedTool: rowReduced.actualTool ?? null
+      });
+    }
+  }
+
+  const tools = [...perTool.values()].map((bucket) => {
+    const fullAccuracy = ratio(bucket.fullCorrect, bucket.total);
+    const reducedAccuracy = ratio(bucket.reducedCorrect, bucket.total);
+    return {
+      ...bucket,
+      fullAccuracy,
+      reducedAccuracy,
+      delta: fullAccuracy === null || reducedAccuracy === null ? null : reducedAccuracy - fullAccuracy
+    };
+  });
+
+  return {
+    model,
+    compared: comparedIds.length,
+    improved,
+    worsened,
+    tools,
+    verdict:
+      improved.length > worsened.length
+        ? "interference"
+        : worsened.length > improved.length
+          ? "helpful"
+          : "neutral"
+  };
+}
+
+export function formatAblation(payload) {
+  const lines = [
+    `ablate: ${payload.tool}（suite ${payload.suite}，去掉后期望该工具的用例 ${payload.droppedCases} 条）`
+  ];
+
+  for (const analysis of payload.models) {
+    lines.push("", `[${analysis.model}] 比较用例 ${analysis.compared} 条`);
+    lines.push("  tool                    full   reduced   delta");
+    lines.push("  ----------------------  -----  --------  -----");
+    for (const bucket of analysis.tools) {
+      lines.push(
+        `  ${bucket.tool.padEnd(22)}  ${percent(bucket.fullAccuracy).padStart(5)}  ${percent(bucket.reducedAccuracy).padStart(8)}  ${signed(bucket.delta).padStart(5)}`
+      );
+    }
+    if (analysis.improved.length > 0) {
+      lines.push("  去掉后变对的用例:");
+      for (const item of analysis.improved) {
+        lines.push(
+          `    - ${item.id}: ${item.fullTool ?? "(none)"} → ${item.reducedTool ?? "(none)"}（期望 ${item.expected ?? "(none)"}）`
+        );
+      }
+    }
+    if (analysis.worsened.length > 0) {
+      lines.push("  去掉后变错的用例:");
+      for (const item of analysis.worsened) {
+        lines.push(
+          `    - ${item.id}: ${item.fullTool ?? "(none)"} → ${item.reducedTool ?? "(none)"}（期望 ${item.expected ?? "(none)"}）`
+        );
+      }
+    }
+    const verdict =
+      analysis.verdict === "interference"
+        ? `存在干扰：去掉 ${payload.tool} 后改善 ${analysis.improved.length} 条、变差 ${analysis.worsened.length} 条`
+        : analysis.verdict === "helpful"
+          ? `有正向作用：去掉后变差 ${analysis.worsened.length} 条多于改善 ${analysis.improved.length} 条`
+          : `影响中性：改善 ${analysis.improved.length} 条、变差 ${analysis.worsened.length} 条`;
+    lines.push(`  结论: ${verdict}`);
+  }
+
+  return lines.join("\n");
+}
+
+function ratio(numerator, denominator) {
+  return denominator ? numerator / denominator : null;
+}
+
+function percent(value) {
+  return value === null ? "n/a" : `${Math.round(value * 100)}%`;
+}
+
+function signed(value) {
+  if (value === null) return "n/a";
+  const rounded = Math.round(value * 100);
+  return `${rounded > 0 ? "+" : ""}${rounded}%`;
+}

package/lib/adapter-import.js

@@ -0,0 +1,24 @@
+import { resolve } from "node:path";
+import { pathToFileURL } from "node:url";
+import { normalizeTool } from "./scan.js";
+
+export async function loadAdapterTools(filePath) {
+  const abs = resolve(filePath);
+  let mod;
+  try {
+    mod = await import(`${pathToFileURL(abs).href}?t=${Date.now()}`);
+  } catch (error) {
+    throw new Error(`adapter import failed (${abs}): ${error.message}`);
+  }
+
+  const raw = mod.default ?? mod.tools ?? mod.functions;
+  if (!Array.isArray(raw)) {
+    throw new Error(`${abs}: 需要 export default [...] 或 export const tools = [...]`);
+  }
+
+  const tools = raw.map(normalizeTool).filter(Boolean);
+  if (tools.length === 0) {
+    throw new Error(`${abs}: 没有解析出任何工具（需要 name + description + input/parameters）`);
+  }
+  return tools;
+}

package/lib/adapters/mock.js

@@ -0,0 +1,143 @@
+const INTENTS = [
+  { tool: "resume_score", words: ["简历", "匹配", "打分", "评分", "resume", "match"] },
+  { tool: "interview_question", words: ["面试", "提问", "interview"] },
+  { tool: "salary_band", words: ["薪资", "薪水", "工资", "salary", "pay", "多少"] },
+  { tool: "jd_rewrite", words: ["改写", "润色", "重写", "rewrite", "jd"] }
+];
+
+const PARAM_HINTS = {
+  resume: ["简历", "resume"],
+  jd: ["职位描述", "岗位描述", "jd"],
+  role: ["岗位", "职位", "role"],
+  city: ["城市", "city"],
+  level: ["级别", "职级", "level"],
+  years: ["年限", "经验", "years"]
+};
+
+const NAME_WEIGHT = 3;
+const DESCRIPTION_WEIGHT = 1.5;
+const PARAM_WEIGHT = 0.5;
+const INTENT_WEIGHT = 2.5;
+const THRESHOLD = 2.5;
+
+export function createMockAdapter() {
+  return {
+    label: "mock-router-v1",
+    async call({ input, tools }) {
+      const started = performance.now();
+      const picked = pickTool(input, tools.map((tool) => tool.function ?? tool));
+      const toolCalls = picked ? [{ name: picked.name, arguments: fillArgs(picked, input) }] : [];
+      const toolNames = tools.map((tool) => (tool.function ?? tool).name).join(" ");
+      const promptChars = input.length + toolNames.length + JSON.stringify(tools).length;
+      const completionChars = JSON.stringify(toolCalls).length;
+      return {
+        ok: true,
+        toolCalls,
+        usage: {
+          input: Math.ceil(promptChars / 2),
+          output: Math.ceil(completionChars / 2),
+          estimated: true
+        },
+        latencyMs: Math.round(performance.now() - started)
+      };
+    }
+  };
+}
+
+function pickTool(input, tools) {
+  const queryTokens = tokenize(input);
+  let best = null;
+  for (const tool of tools) {
+    const score = scoreTool(tool, queryTokens, input);
+    if (!best || score > best.score) best = { tool, score };
+  }
+  return best && best.score >= THRESHOLD ? best.tool : null;
+}
+
+function scoreTool(tool, queryTokens, input) {
+  const lowered = input.toLowerCase();
+  let score = overlap(queryTokens, tokenize(tool.name)) * NAME_WEIGHT;
+  score += overlap(queryTokens, tokenize(tool.description)) * DESCRIPTION_WEIGHT;
+
+  const props = tool.parameters?.properties ?? {};
+  for (const [name, schema] of Object.entries(props)) {
+    score += overlap(queryTokens, tokenize(`${name} ${schema?.description ?? ""}`)) * PARAM_WEIGHT;
+  }
+
+  for (const intent of INTENTS) {
+    if (intent.tool !== tool.name) continue;
+    if (intent.words.some((word) => lowered.includes(word.toLowerCase()))) {
+      score += INTENT_WEIGHT;
+      break;
+    }
+  }
+  return score;
+}
+
+function overlap(a, b) {
+  if (a.size === 0 || b.size === 0) return 0;
+  let hits = 0;
+  for (const token of b) if (a.has(token)) hits += 1;
+  return hits;
+}
+
+function tokenize(text) {
+  const tokens = new Set();
+  const lowered = String(text ?? "").toLowerCase();
+  for (const word of lowered.match(/[a-z][a-z0-9_]{1,}/g) ?? []) tokens.add(word);
+  for (const chunk of lowered.match(/[一-鿿]+/g) ?? []) {
+    if (chunk.length === 1) tokens.add(chunk);
+    for (let i = 0; i + 1 < chunk.length; i += 1) tokens.add(chunk.slice(i, i + 2));
+  }
+  return tokens;
+}
+
+function fillArgs(tool, input) {
+  const sections = parseSections(input);
+  const args = {};
+  const props = tool.parameters?.properties ?? {};
+  for (const name of Object.keys(props)) {
+    const value = findSectionValue(name, sections);
+    if (value !== undefined) args[name] = value;
+  }
+  return args;
+}
+
+function parseSections(text) {
+  const lines = String(text).split(/\r?\n/);
+  const sections = [];
+  for (let i = 0; i < lines.length; i += 1) {
+    const match = lines[i].match(/^\s*([^\s：:]{1,20})\s*[：:]\s*(.*)$/);
+    if (!match) continue;
+    const header = match[1].trim();
+    let value = match[2].trim();
+    if (value === "") {
+      const body = [];
+      let j = i + 1;
+      while (j < lines.length) {
+        const line = lines[j].trim();
+        if (line === "" || /^[^\s：:]{1,20}\s*[：:]/.test(line)) break;
+        body.push(line);
+        j += 1;
+      }
+      value = body.join("\n");
+    }
+    if (value !== "") sections.push({ header, value });
+  }
+  return sections;
+}
+
+function findSectionValue(paramName, sections) {
+  const hints = PARAM_HINTS[paramName] ?? [paramName];
+  const loweredHints = hints.map((hint) => hint.toLowerCase());
+  const hit = sections.find((section) => {
+    const header = section.header.toLowerCase();
+    return loweredHints.some((hint) => header.includes(hint));
+  });
+  if (!hit) return undefined;
+  if (paramName === "years") {
+    const number = hit.value.match(/\d+(?:\.\d+)?/);
+    return number ? Number(number[0]) : undefined;
+  }
+  return hit.value;
+}