Files
thzxx 0fc589e73b feat: 升级至 v0.3.2 — 工具体系扩展至 26 个 + Context Window 可配置化
## 主要变更

### 1. Context Window 可配置化(v0.3.1 延续)
- DeepSeek/Agnes contextWindow 不再写死 1M,可在设置中配置(min 4096)
- 修复 Engine 128K vs Adapter 1M 不一致 bug,Engine 从 adapter.getContextWindow() 读取
- 热重载 configSig 加入 contextWindow,配置变化即时生效

### 2. 新增 11 个工具(15 → 26 个)
- Git 工具集(4):git_status, git_diff, git_log, git_commit
- 开发工具集(3):lint_code, run_tests, project_info
- HTTP 请求(1):http_request(Node 18+ fetch + AbortController)
- TODO 管理(1):todo_write(会话级内存 + LRU 淘汰)
- 结构化思考(1):think(无副作用思考空间)
- 图片查看(1):view_image(base64 data URL,多模态 LLM 支持)

### 3. 审计修复(2 FAIL + 14 WARN)
- FAIL-1: registry.ts truncateResult 添加 dataUrl 白名单(图片不被截断)
- FAIL-2: todo.ts 添加 LRU 策略 + clearSession 静态方法
- WARN-1: run_tests filter 字符白名单校验,防 cmd 元字符注入
- WARN-2: dataUrl 检测前置到 stringify 之前,避免大图片无意义序列化
- WARN-3: todo.ts 实现真正 LRU(访问刷新位置,非 FIFO)
- WARN-4: git_diff maxBuffer 提升至 5MB,支持超大变更集
- WARN-5: log.info 移至 DelegateTaskTool 注册后,输出正确的 26
- WARN-6: riskColors 添加 critical: 'error' 键
- WARN-7: 所有 execFileAsync 显式设置 encoding: 'utf-8'
- WARN-8: git_log --author 拆分为独立参数
- WARN-9: todo_write 权限从 WRITE 改为 READ
- WARN-10: description 区分与 task_manager 的不同用途

### 4. PolicyEngine 策略
- 新增 11 条策略,26 个工具 + mcp_* 全覆盖
- git_commit: WRITE + requireConfirmation
- todo_write: READ(内存操作)
2026-07-14 21:15:54 +08:00

52 lines
1.9 KiB
TypeScript

/**
* 思考工具(1 个)
*
* think — 结构化思考空间(无副作用)
*
* 参考 Claude Code 的 think 工具设计。
* 帮助 LLM 在工具调用上下文中组织思路。
*/
import type { IMetonaTool, ToolExecutionContext } from '../../types/metona-tool';
import type { MetonaToolDef } from '../../../harness/types';
import { MetonaToolCategory, MetonaRiskLevel } from '../../../harness/types';
export class ThinkTool implements IMetonaTool {
readonly definition: MetonaToolDef = {
name: 'think',
description:
'Use this tool for structured thinking. It helps you organize your thoughts before complex actions. This tool has no side effects — it simply returns your input for your own reference. Use it when: 1) Breaking down complex tasks into steps, 2) Evaluating multiple approaches, 3) Planning before writing code, 4) Reviewing your work before completing.',
parameters: {
type: 'object',
properties: {
thought: { type: 'string', description: 'The thought content' },
nextAction: { type: 'string', description: 'Next action plan' },
reasoning: { type: 'string', description: 'Reasoning process' },
},
required: ['thought'],
},
category: MetonaToolCategory.CUSTOM,
riskLevel: MetonaRiskLevel.SAFE,
requiresPermission: false,
timeoutMs: 1_000,
};
async execute(args: Record<string, unknown>, _context: ToolExecutionContext): Promise<unknown> {
try {
const thought = args.thought as string;
const nextAction = (args.nextAction as string) ?? null;
const reasoning = (args.reasoning as string) ?? null;
return {
thought,
nextAction,
reasoning,
timestamp: Date.now(),
};
} catch (error) {
const errMsg = error instanceof Error ? error.message : String(error);
return { error: errMsg, success: false };
}
}
}