lc_tools/
url_fetch.rs

1// lc-tools/src/url_fetch.rs
2//! Web page fetching tool
3//!
4//! Provides web content fetching and parsing.
5
6use async_trait::async_trait;
7use regex::Regex;
8use schemars::JsonSchema;
9use serde::{Deserialize, Serialize};
10
11use crate::ssrf::guarded_get;
12use lc_core::tools::{BaseTool, Tool, ToolError};
13
14static SCRIPT_REGEX: std::sync::LazyLock<Regex> =
15    std::sync::LazyLock::new(|| Regex::new(r"<script[^>]*>.*?</script>").unwrap());
16
17static STYLE_REGEX: std::sync::LazyLock<Regex> =
18    std::sync::LazyLock::new(|| Regex::new(r"<style[^>]*>.*?</style>").unwrap());
19
20static TAG_REGEX: std::sync::LazyLock<Regex> =
21    std::sync::LazyLock::new(|| Regex::new(r"<[^>]+>").unwrap());
22
23static WHITESPACE_REGEX: std::sync::LazyLock<Regex> =
24    std::sync::LazyLock::new(|| Regex::new(r"\s+").unwrap());
25
26static LINK_REGEX: std::sync::LazyLock<Regex> =
27    std::sync::LazyLock::new(|| Regex::new(r#"<a[^>]+href\s*=\s*['"]([^'"]+)['"][^>]*>"#).unwrap());
28
29static IMG_REGEX: std::sync::LazyLock<Regex> = std::sync::LazyLock::new(|| {
30    Regex::new(r#"<img[^>]+src\s*=\s*['"]([^'"]+)['"][^>]*>"#).unwrap()
31});
32
33static TITLE_REGEX: std::sync::LazyLock<Regex> =
34    std::sync::LazyLock::new(|| Regex::new(r"<title[^>]*>(.*?)</title>").unwrap());
35
36static DESC_REGEX: std::sync::LazyLock<Regex> = std::sync::LazyLock::new(|| {
37    Regex::new(r#"<meta[^>]+name\s*=\s*['"]description['"][^>]+content\s*=\s*['"]([^'"]+)['"]"#)
38        .unwrap()
39});
40
41static KW_REGEX: std::sync::LazyLock<Regex> = std::sync::LazyLock::new(|| {
42    Regex::new(r#"<meta[^>]+name\s*=\s*['"]keywords['"][^>]+content\s*=\s*['"]([^'"]+)['"]"#)
43        .unwrap()
44});
45
46/// Extracts links from HTML and dedups them (preserving first-occurrence order).
47///
48/// Returns `(deduped links, raw count)`. The raw count is used by details to show
49/// "before/after dedup". Q4: the old implementation only renamed the Vec to
50/// `unique_links` without actually deduplicating.
51fn extract_unique_links(html: &str) -> (Vec<String>, usize) {
52    let raw: Vec<String> = LINK_REGEX
53        .captures_iter(html)
54        .map(|cap| cap[1].to_string())
55        .collect();
56    let raw_count = raw.len();
57    let mut seen = std::collections::HashSet::new();
58    let unique: Vec<String> = raw
59        .into_iter()
60        .filter(|link| seen.insert(link.clone()))
61        .collect();
62    (unique, raw_count)
63}
64
65/// URLFetch tool input
66#[derive(Debug, Deserialize, JsonSchema)]
67pub struct URLFetchInput {
68    /// Operation type: "fetch", "extract_text", "extract_links", "extract_images", "metadata"
69    pub operation: String,
70
71    /// The URL
72    pub url: String,
73
74    /// Whether to include header info (for the fetch operation)
75    pub include_headers: Option<bool>,
76
77    /// Maximum content length (bytes)
78    pub max_length: Option<usize>,
79}
80
81/// URLFetch tool output
82#[derive(Debug, Serialize)]
83pub struct URLFetchOutput {
84    /// Operation result
85    pub result: String,
86
87    /// Operation type
88    pub operation: String,
89
90    /// URL
91    pub url: String,
92
93    /// Content length
94    pub content_length: usize,
95
96    /// Extra details
97    pub details: Option<String>,
98}
99
100/// Web page fetching tool
101pub struct URLFetchTool {
102    /// Whether access to private/internal IPs is allowed (default false)
103    allow_private_ips: bool,
104}
105
106impl URLFetchTool {
107    /// Creates a web fetching tool (SSRF protection enabled by default).
108    ///
109    /// The HTTP client is built per request inside `guarded_get` so the
110    /// validated DNS answers can be pinned onto it (A3 DNS-rebinding closure).
111    pub fn new() -> Self {
112        Self {
113            allow_private_ips: false,
114        }
115    }
116
117    /// Allow requests to private/internal IP addresses (SSRF opt-in).
118    pub fn with_allow_private_ips(mut self, allow: bool) -> Self {
119        self.allow_private_ips = allow;
120        self
121    }
122
123    /// Fetches web content
124    async fn fetch_url(
125        &self,
126        url: &str,
127        max_length: Option<usize>,
128        include_headers: Option<bool>,
129    ) -> Result<URLFetchOutput, ToolError> {
130        if !url.starts_with("http://") && !url.starts_with("https://") {
131            return Err(ToolError::InvalidInput(
132                "URL must start with http:// or https://".to_string(),
133            ));
134        }
135
136        // SSRF: guarded_get checks each hop, pins the validated IPs, and follows
137        // redirects manually (both the first hop and every redirect target are
138        // re-resolved and re-checked against intranet addresses).
139        let response = guarded_get(
140            url,
141            !self.allow_private_ips,
142            Some(std::time::Duration::from_secs(30)),
143        )
144        .await?;
145
146        let status = response.status();
147        if !status.is_success() {
148            return Err(ToolError::ExecutionFailed(format!(
149                "HTTP error: {} - {}",
150                status.as_u16(),
151                status.canonical_reason().unwrap_or("unknown")
152            )));
153        }
154
155        // Q3: when include_headers = Some(true), merge the response headers into the output
156        // (details) instead of silently ignoring them. Headers must be read before consuming
157        // the response (reqwest's headers() borrows, text() consumes).
158        let header_block: String = if include_headers.unwrap_or(false) {
159            let mut lines: Vec<String> = response
160                .headers()
161                .iter()
162                .map(|(name, value)| format!("{}: {}", name, value.to_str().unwrap_or("<非UTF-8>")))
163                .collect();
164            lines.sort();
165            let mut block = String::from("响应头:\n");
166            for line in lines {
167                block.push_str(&line);
168                block.push('\n');
169            }
170            block
171        } else {
172            String::new()
173        };
174
175        let content = response
176            .text()
177            .await
178            .map_err(|e| ToolError::ExecutionFailed(format!("failed to read response: {}", e)))?;
179
180        let max_len = max_length.unwrap_or(50000);
181        let content_len = content.len();
182        let truncated = content_len > max_len;
183        let result = if truncated {
184            content.chars().take(max_len).collect::<String>() + "\n... [内容已截断]"
185        } else {
186            content
187        };
188
189        let mut details = format!(
190            "状态码: {}, 内容长度: {} 字节{}",
191            status.as_u16(),
192            content_len,
193            if truncated { " (已截断)" } else { "" }
194        );
195        if !header_block.is_empty() {
196            details.push('\n');
197            details.push_str(&header_block);
198        }
199
200        Ok(URLFetchOutput {
201            result,
202            operation: "fetch".to_string(),
203            url: url.to_string(),
204            content_length: content_len,
205            details: Some(details),
206        })
207    }
208
209    /// Extracts plain text content
210    async fn extract_text(&self, url: &str) -> Result<URLFetchOutput, ToolError> {
211        let fetch_result = self.fetch_url(url, Some(100000), None).await?;
212        let html = &fetch_result.result;
213
214        let html = SCRIPT_REGEX.replace_all(html, "");
215        let html = STYLE_REGEX.replace_all(&html, "");
216
217        let text = TAG_REGEX.replace_all(&html, "");
218
219        let clean_text = WHITESPACE_REGEX.replace_all(&text, " ").trim().to_string();
220
221        let max_len = 5000;
222        let clean_len = clean_text.len();
223        let result = if clean_len > max_len {
224            clean_text.chars().take(max_len).collect::<String>() + "..."
225        } else {
226            clean_text
227        };
228
229        Ok(URLFetchOutput {
230            result,
231            operation: "extract_text".to_string(),
232            url: url.to_string(),
233            content_length: clean_len,
234            details: Some(format!("提取了 {} 字符的纯文本", clean_len)),
235        })
236    }
237
238    /// Extracts links (deduped, first-occurrence order preserved)
239    async fn extract_links(&self, url: &str) -> Result<URLFetchOutput, ToolError> {
240        let fetch_result = self.fetch_url(url, Some(100000), None).await?;
241        let html = &fetch_result.result;
242
243        let (unique_links, raw_count) = extract_unique_links(html);
244        let result = unique_links.join("\n");
245
246        Ok(URLFetchOutput {
247            result,
248            operation: "extract_links".to_string(),
249            url: url.to_string(),
250            content_length: html.len(), // Q4: the real body length, not the link count
251            details: Some(format!(
252                "找到 {} 个唯一链接(原始 {} 个)",
253                unique_links.len(),
254                raw_count
255            )),
256        })
257    }
258
259    /// Extracts image links
260    async fn extract_images(&self, url: &str) -> Result<URLFetchOutput, ToolError> {
261        let fetch_result = self.fetch_url(url, Some(100000), None).await?;
262        let html = &fetch_result.result;
263
264        let images: Vec<String> = IMG_REGEX
265            .captures_iter(html)
266            .map(|cap| cap[1].to_string())
267            .collect();
268
269        let result = images.join("\n");
270
271        Ok(URLFetchOutput {
272            result,
273            operation: "extract_images".to_string(),
274            url: url.to_string(),
275            content_length: images.len(),
276            details: Some(format!("找到 {} 张图片", images.len())),
277        })
278    }
279
280    /// Extracts metadata
281    async fn extract_metadata(&self, url: &str) -> Result<URLFetchOutput, ToolError> {
282        let fetch_result = self.fetch_url(url, Some(50000), None).await?;
283        let html = &fetch_result.result;
284
285        let title = TITLE_REGEX
286            .captures(html)
287            .map(|cap| cap[1].trim().to_string())
288            .unwrap_or_default();
289
290        let description = DESC_REGEX
291            .captures(html)
292            .map(|cap| cap[1].to_string())
293            .unwrap_or_default();
294
295        let keywords = KW_REGEX
296            .captures(html)
297            .map(|cap| cap[1].to_string())
298            .unwrap_or_default();
299
300        let result = format!(
301            "标题: {}\n描述: {}\n关键词: {}",
302            title, description, keywords
303        );
304
305        Ok(URLFetchOutput {
306            result,
307            operation: "metadata".to_string(),
308            url: url.to_string(),
309            content_length: title.len() + description.len() + keywords.len(),
310            details: Some("提取了网页元数据".to_string()),
311        })
312    }
313}
314
315impl Default for URLFetchTool {
316    fn default() -> Self {
317        Self::new()
318    }
319}
320
321#[async_trait]
322impl Tool for URLFetchTool {
323    type Input = URLFetchInput;
324    type Output = URLFetchOutput;
325
326    async fn invoke(&self, input: Self::Input) -> Result<Self::Output, ToolError> {
327        match input.operation.as_str() {
328            "fetch" => {
329                self.fetch_url(&input.url, input.max_length, input.include_headers)
330                    .await
331            }
332            "extract_text" => self.extract_text(&input.url).await,
333            "extract_links" => self.extract_links(&input.url).await,
334            "extract_images" => self.extract_images(&input.url).await,
335            "metadata" => self.extract_metadata(&input.url).await,
336            _ => Err(ToolError::InvalidInput(
337                format!("unsupported operation: {}, use: fetch, extract_text, extract_links, extract_images, metadata", input.operation)
338            )),
339        }
340    }
341}
342
343#[async_trait]
344impl BaseTool for URLFetchTool {
345    fn name(&self) -> &str {
346        "url_fetch"
347    }
348
349    fn description(&self) -> &str {
350        "网页抓取工具。支持多种操作：
351
352操作类型:
353- fetch: 抓取完整网页内容
354- extract_text: 提取纯文本内容（去除HTML标签）
355- extract_links: 提取所有链接
356- extract_images: 提取所有图片链接
357- metadata: 提取网页元数据（标题、描述、关键词）
358
359参数:
360- url: 网页地址（必须以 http:// 或 https:// 开头）
361- max_length: 最大内容长度（可选，默认50KB）
362- include_headers: 是否包含头部信息（可选）
363
364示例:
365- 抓取网页: {\"operation\": \"fetch\", \"url\": \"https://example.com\"}
366- 提取文本: {\"operation\": \"extract_text\", \"url\": \"https://example.com\"}
367- 提取链接: {\"operation\": \"extract_links\", \"url\": \"https://example.com\"}"
368    }
369
370    async fn run(&self, input: String) -> Result<String, ToolError> {
371        let parsed: URLFetchInput = serde_json::from_str(&input)
372            .map_err(|e| ToolError::InvalidInput(format!("JSON parse failed: {}", e)))?;
373
374        let output = self.invoke(parsed).await?;
375
376        Ok(format!(
377            "URL: {}\n操作: {}\n内容长度: {} 字节\n\n{}\n详细信息: {}",
378            output.url,
379            output.operation,
380            output.content_length,
381            output.result,
382            output.details.unwrap_or_default()
383        ))
384    }
385
386    fn args_schema(&self) -> Option<serde_json::Value> {
387        use schemars::schema_for;
388        serde_json::to_value(schema_for!(URLFetchInput)).ok()
389    }
390}
391
392#[cfg(test)]
393mod tests {
394    use super::*;
395
396    #[test]
397    fn test_url_validation() {
398        let valid_url = "https://example.com";
399        assert!(valid_url.starts_with("http://") || valid_url.starts_with("https://"));
400
401        let valid_url2 = "http://example.org";
402        assert!(valid_url2.starts_with("http://") || valid_url2.starts_with("https://"));
403    }
404
405    #[tokio::test]
406    async fn test_url_fetch_invalid_url() {
407        let tool = URLFetchTool::new();
408
409        let input = URLFetchInput {
410            operation: "fetch".to_string(),
411            url: "invalid-url".to_string(),
412            include_headers: None,
413            max_length: None,
414        };
415
416        let result = tool.invoke(input).await;
417        assert!(result.is_err());
418        assert!(result.unwrap_err().to_string().contains("http://"));
419    }
420
421    /// Q4: extract_links truly dedups and preserves first-occurrence order; content_length is the body length.
422    #[test]
423    fn test_extract_unique_links_dedups() {
424        let html = r#"
425            <a href="https://a.com/1">first</a>
426            <a href="https://a.com/1">dup</a>
427            <a href="https://a.com/2">second</a>
428            <a href="https://a.com/1">dup2</a>
429        "#;
430        let (unique, raw) = extract_unique_links(html);
431        assert_eq!(raw, 4, "原始链接数应为 4");
432        assert_eq!(
433            unique,
434            vec!["https://a.com/1".to_string(), "https://a.com/2".to_string()],
435            "应去重且保持首次出现顺序"
436        );
437    }
438
439    /// Q1: after SSRF was extracted into a shared module, URLFetch still blocks intranet addresses by default.
440    #[tokio::test]
441    async fn test_url_fetch_blocks_localhost_by_default() {
442        let tool = URLFetchTool::new();
443        let input = URLFetchInput {
444            operation: "fetch".to_string(),
445            url: "http://127.0.0.1:6379/".to_string(),
446            include_headers: None,
447            max_length: None,
448        };
449        let result = tool.invoke(input).await;
450        assert!(result.is_err());
451        let err = result.unwrap_err().to_string();
452        assert!(err.contains("SSRF"), "expected SSRF error, got: {}", err);
453    }
454
455    /// Q3: when include_headers = Some(true), response headers are merged into the output (details).
456    #[tokio::test]
457    async fn test_fetch_include_headers() {
458        use tokio::io::{AsyncReadExt, AsyncWriteExt};
459        use tokio::net::TcpListener;
460
461        let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
462        let addr = listener.local_addr().unwrap();
463
464        let server = tokio::spawn(async move {
465            let (mut socket, _) = listener.accept().await.unwrap();
466            let mut buf = [0u8; 4096];
467            let _ = socket.read(&mut buf).await;
468            let body = "hello world";
469            let resp = format!(
470                "HTTP/1.1 200 OK\r\nContent-Length: {}\r\nContent-Type: text/plain\r\nX-Test-Header: hello\r\n\r\n{}",
471                body.len(),
472                body
473            );
474            socket.write_all(resp.as_bytes()).await.unwrap();
475            drop(socket);
476        });
477
478        let tool = URLFetchTool::new().with_allow_private_ips(true);
479        let input = URLFetchInput {
480            operation: "fetch".to_string(),
481            url: format!("http://{}/", addr),
482            include_headers: Some(true),
483            max_length: None,
484        };
485
486        let output = tool.invoke(input).await.unwrap();
487        let details = output.details.unwrap();
488        assert!(details.contains("响应头:"), "details: {}", details);
489        // reqwest normalizes response header names to lowercase (HTTP header names are case-insensitive)
490        assert!(
491            details.contains("x-test-header: hello"),
492            "details: {}",
493            details
494        );
495        assert!(output.result.contains("hello world"));
496        server.await.unwrap();
497    }
498
499    /// Q3: when include_headers = false/None, no response headers are returned.
500    #[tokio::test]
501    async fn test_fetch_without_include_headers() {
502        use tokio::io::{AsyncReadExt, AsyncWriteExt};
503        use tokio::net::TcpListener;
504
505        let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
506        let addr = listener.local_addr().unwrap();
507
508        let server = tokio::spawn(async move {
509            let (mut socket, _) = listener.accept().await.unwrap();
510            let mut buf = [0u8; 4096];
511            let _ = socket.read(&mut buf).await;
512            let body = "hello world";
513            let resp = format!(
514                "HTTP/1.1 200 OK\r\nContent-Length: {}\r\nX-Test-Header: hello\r\n\r\n{}",
515                body.len(),
516                body
517            );
518            socket.write_all(resp.as_bytes()).await.unwrap();
519            drop(socket);
520        });
521
522        let tool = URLFetchTool::new().with_allow_private_ips(true);
523        let input = URLFetchInput {
524            operation: "fetch".to_string(),
525            url: format!("http://{}/", addr),
526            include_headers: Some(false),
527            max_length: None,
528        };
529
530        let output = tool.invoke(input).await.unwrap();
531        let details = output.details.unwrap();
532        assert!(
533            !details.contains("X-Test-Header"),
534            "不应包含响应头, got: {}",
535            details
536        );
537        server.await.unwrap();
538    }
539
540    #[test]
541    fn test_tool_properties() {
542        let tool = URLFetchTool::new();
543
544        assert_eq!(tool.name(), "url_fetch");
545        assert!(tool.description().contains("fetch"));
546        assert!(BaseTool::args_schema(&tool).is_some());
547    }
548}
lc_tools/url_fetch.rs

lc_tools/
url_fetch.rs
