tool-selector

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Tool Selector Protocol

工具选择器协议

Determine which tools are needed to solve the task, plan the most efficient sequence, and eliminate unnecessary calls before they happen.

Core principle: Every tool call costs time, tokens, and rate limits. The right tool, in the right order, called once.

确定解决任务所需的工具，规划最高效的执行顺序，并提前避免不必要的工具调用。

核心原则：每次工具调用都会消耗时间、令牌（tokens）和调用限额。要选择正确的工具，按正确的顺序，只调用一次。

Tool Catalog

工具目录

Tool	Cost	Side Effect	Best For
Web search	Medium	None (read-only)	Current info, external sources
Code execution	Medium	Environment-dependent	Computation, transformation, testing
File read	Low	None	Local data, config, content
File write	Medium	Persistent change	Saving output, updates
API call	High	Rate limits, cost	External service integration
Database query	Medium-High	Persistent on writes	Data read/write
Memory/context	Low	None	Retrieve previous information

工具	成本	副作用	最佳适用场景
网页搜索	中等	无（只读）	最新信息、外部资源
代码执行	中等	取决于运行环境	计算、转换、测试
文件读取	低	无	本地数据、配置、内容
文件写入	中等	产生持久化变更	保存输出、更新内容
API调用	高	受调用限额、成本限制	外部服务集成
数据库查询	中-高	写入操作会产生持久化变更	数据读写
内存/上下文	低	无	检索历史信息

Selection Algorithm

选择算法

For each tool candidate, ask these questions in order:

1. Can this tool actually solve this task?           → No = eliminate
2. Can a cheaper/simpler tool do the same job?       → Yes = use that instead
3. Was this tool already called for this task?        → Yes = use cached result
4. Is this call actually needed right now?            → No = defer
5. Does this tool depend on another tool's output?    → Yes = add to sequence

对于每个候选工具，按顺序提出以下问题：

1. Can this tool actually solve this task?           → No = eliminate
2. Can a cheaper/simpler tool do the same job?       → Yes = use that instead
3. Was this tool already called for this task?        → Yes = use cached result
4. Is this call actually needed right now?            → No = defer
5. Does this tool depend on another tool's output?    → Yes = add to sequence

Anti-Patterns — Block These

反模式 — 需避免

Anti-Pattern	Problem	Solution
Reading the same file repeatedly	Each read costs tokens	Cache on first read
Web search for known information	Unnecessary latency	Use existing knowledge
Serial API calls (could be parallel)	Slow	Plan parallel calls
Tool chain when one tool suffices	Extra steps	Consolidate
Read before write (when unnecessary)	Extra call	Skip the read
Loading memory every step	Token waste	Load once, reference

反模式	问题	解决方案
重复读取同一文件	每次读取都会消耗令牌	首次读取后缓存结果
对已知信息进行网页搜索	造成不必要的延迟	使用已有知识
串行API调用（可并行）	速度慢	规划并行调用
可用单个工具却使用工具链	额外步骤	合并操作
不必要的先读后写	额外调用	跳过读取步骤
每步都加载内存	浪费令牌	加载一次后复用

Output Format

输出格式

TOOL SELECTOR
Task    : [task summary]
Selected: N tools | Total calls: N | Blocked: N

TOOL SELECTOR
Task    : [task summary]
Selected: N tools | Total calls: N | Blocked: N

Tool Plan

Step	Tool	Purpose	Depends On	Cached?
1	[tool]	[why]	—	No
2	[tool]	[why]	#1	Yes — use #1 output

Step	Tool	Purpose	Depends On	Cached?
1	[tool]	[why]	—	No
2	[tool]	[why]	#1	Yes — use #1 output

Blocked Calls

Blocked	Reason	Alternative
[tool call]	[rationale]	[what to do instead]

Blocked	Reason	Alternative
[tool call]	[rationale]	[what to do instead]

Optimization Notes

[Parallel opportunities]
[Cached outputs]
[Rate limit warnings]

Approve? Proceed?

---

[Parallel opportunities]
[Cached outputs]
[Rate limit warnings]

Approve? Proceed?

---

Tool Combination Guide

工具组合指南

Task Type	Typical Chain	Block
Research & Report	Search → File read → Code (analyze) → File write	Searching same topic twice
Data Processing	File read → Code → File write	Re-reading file each step
API Integration	Memory (creds) → API → Code (parse) → DB write	Re-auth per request
Code Development	File read → Code (test) → File write → Code (verify)	Full test suite on every change
Database Operation	Memory (schema) → DB (SELECT) → Code (prepare) → DB (write)	Running same SELECT twice

任务类型	典型工具链	需避免的操作
研究与报告	网页搜索 → 文件读取 → 代码（分析）→ 文件写入	重复搜索同一主题
数据处理	文件读取 → 代码执行 → 文件写入	每步重复读取文件
API集成	内存（凭证）→ API调用 → 代码（解析）→ 数据库写入	每次请求重复认证
代码开发	文件读取 → 代码（测试）→ 文件写入 → 代码（验证）	每次变更都运行完整测试套件
数据库操作	内存（Schema）→ 数据库（查询）→ 代码（预处理）→ 数据库（写入）	重复执行同一查询

When to Skip

跳过场景

Single-tool task ("read this file")
User already specified which tool to use
Single-step, side-effect-free task

单工具任务（如“读取此文件”）
用户已指定要使用的工具
单步骤、无副作用的任务

Guardrails

防护规则

Never call an API when cached data suffices — always check cache first.
Prefer low-cost tools — file read over web search when local data exists.
Cross-skill: works with
```
parallel-planner
```
(parallel tool calls) and
```
task-decomposer
```
(tool needs per subtask).

当缓存数据可用时，绝不调用API — 始终先检查缓存。
优先使用低成本工具 — 当存在本地数据时，优先使用文件读取而非网页搜索。
跨技能兼容：可与
```
parallel-planner
```
（并行工具调用）和
```
task-decomposer
```
（子任务工具需求）配合使用。