[Bug]: 都不能操作文件什么的，这就是问答的吗？

[gxlqssjf]

Bug type

Regression (worked before, now fails)

Summary

都不能操作文件什么的，这就是问答的吗？说好的个人助理呢？

Steps to reproduce

都不能操作文件什么的，这就是问答的吗？说好的个人助理呢？

Expected behavior

都不能操作文件什么的，这就是问答的吗？说好的个人助理呢？

Actual behavior

都不能操作文件什么的，这就是问答的吗？说好的个人助理呢？

OpenClaw version

2026.03.28

Operating system

ubuntu 24.04

Install method

No response

Model

所有

Provider / routing chain

聊天

Config file / key location

No response

Additional provider/model setup details

都不能操作文件什么的，这就是问答的吗？说好的个人助理呢？

Logs, screenshots, and evidence

都不能操作文件什么的，这就是问答的吗？说好的个人助理呢？

Impact and severity

都不能操作文件什么的，这就是问答的吗？说好的个人助理呢？

Additional information

都不能操作文件什么的，这就是问答的吗？说好的个人助理呢？

[sharkfee]

修改~\openclaw-zero-token\src\zero-token\tool-calling 下面的 web-tool-prompt.ts 的文件，然后重新编译，就可以了。代码里的这句：读取 ~/.openclaw/workspace/AGENTS.md,按着配置文件操作。路径更具你的实际情况该就可以了。 /new 后没有读取配置文件，你输入继续就可以了

/**

• Per-model tool calling prompt templates.
• Reference:
• • Paper: https://arxiv.org/html/2407.04997v1
• • ComfyUI LLM Party: https://github.com/heshengtao/comfyui_LLM_party
    */

import { toolDefsJson } from "./web-tool-defs.js";

const TOOL_DEFS = toolDefsJson();

// Example-based teaching (key insight from arXiv:2407.04997 and ComfyUI LLM Party):
// A trivial example teaches the model the output format without confusing it with real tools.
const TOOL_EXAMPLE = ## Tool Call Format You MUST use EXACT XML format to call tools. Plain text descriptions WILL NOT work. Single tool call example: <tool_call id="call_001" name="plus_one"> {"name":"plus_one","arguments":{"number":"5"}} </tool_call> (plus_one is just an example, not a real tool);

const EN_TEMPLATE = `Tools: ${TOOL_DEFS}

${TOOL_EXAMPLE}

Your actual tools are listed above. To use one, reply ONLY with the XML tool_call block,
and ensure the inner JSON includes "name" and "arguments".
No tool needed? Answer directly.

`;

const EN_STRICT_TEMPLATE = `Tools: ${TOOL_DEFS}

${TOOL_EXAMPLE}

Your actual tools are listed above. To use one, reply ONLY with the XML tool_call block. No extra text.
No tool needed? Answer directly.

`;

const CN_TEMPLATE = `工具: ${TOOL_DEFS}

工具调用格式

必须使用精确的XML格式调用工具，纯文本描述无效。
单工具调用示例：
<tool_call id="call_001" name="plus_one">
{"name":"plus_one","arguments":{"number":"5"}}
</tool_call>
(plus_one仅为示例，非真实工具)

你的真实工具见上方列表。需要时只回复XML tool_call块，
且块内JSON必须包含"name"和"arguments"。不需要则直接回答。
读取 ~/.openclaw/workspace/AGENTS.md,按着配置文件操作。
`;

/** No web models skip prompt injection — web interfaces don't pass native tools.

• Even DeepSeek/Claude/GLM need prompt injection when accessed via browser. */
  const NATIVE_TOOL_MODELS = new Set();

/** Models excluded from tool calling entirely */
const EXCLUDED_MODELS = new Set(["perplexity-web", "doubao-web"]);

/** Chinese-language models */
const CN_MODELS = new Set([
"deepseek-web",
"doubao-web",
"qwen-cn-web",
"kimi-web",
"glm-web",
"xiaomimo-web",
]);

/** Models that tend to add extra text after XML */
const STRICT_MODELS = new Set(["chatgpt-web"]);

export function shouldInjectToolPrompt(api: string): boolean {
return !NATIVE_TOOL_MODELS.has(api) && !EXCLUDED_MODELS.has(api);
}

export function getToolPrompt(api: string): string {
if (STRICT_MODELS.has(api)) {
return EN_STRICT_TEMPLATE;
}
if (CN_MODELS.has(api)) {
return CN_TEMPLATE;
}
return EN_TEMPLATE;
}

/** Format tool result for feedback to the model */
export function formatToolResult(toolName: string, result: string): string {
return Tool ${toolName} returned: ${result}\nPlease continue answering based on this result.;
}