1use agent_settings::AgentProfileId;
2use anyhow::Result;
3use async_trait::async_trait;
4use markdown::PathWithRange;
5
6use crate::example::{Example, ExampleContext, ExampleMetadata, JudgeAssertion, LanguageServer};
7
8pub struct CodeBlockCitations;
9
10const FENCE: &str = "```";
11
12#[async_trait(?Send)]
13impl Example for CodeBlockCitations {
14 fn meta(&self) -> ExampleMetadata {
15 ExampleMetadata {
16 name: "code_block_citations".to_string(),
17 url: "https://github.com/zed-industries/zed.git".to_string(),
18 revision: "f69aeb6311dde3c0b8979c293d019d66498d54f2".to_string(),
19 language_server: Some(LanguageServer {
20 file_extension: "rs".to_string(),
21 allow_preexisting_diagnostics: false,
22 }),
23 max_assertions: None,
24 profile_id: AgentProfileId::default(),
25 existing_thread_json: None,
26 max_turns: None,
27 }
28 }
29
30 async fn conversation(&self, cx: &mut ExampleContext) -> Result<()> {
31 const FILENAME: &str = "assistant_tool.rs";
32
33 // Verify that the messages all have the correct formatting.
34 let texts: Vec<String> = cx
35 .prompt(format!(
36 r#"
37 Show me the method bodies of all the methods of the `Tool` trait in {FILENAME}.
38
39 Please show each method in a separate code snippet.
40 "#
41 ))
42 .await?
43 .texts()
44 .collect();
45 let closing_fence = format!("\n{FENCE}");
46
47 for text in texts.iter() {
48 let mut text = text.as_str();
49
50 while let Some(index) = text.find(FENCE) {
51 // Advance text past the opening backticks.
52 text = &text[index + FENCE.len()..];
53
54 // Find the closing backticks.
55 let content_len = text.find(&closing_fence);
56
57 // Verify the citation format - e.g. ```path/to/foo.txt#L123-456
58 if let Some(citation_len) = text.find('\n') {
59 let citation = &text[..citation_len];
60
61 if let Ok(()) =
62 cx.assert(citation.contains("/"), format!("Slash in {citation:?}",))
63 {
64 let path_range = PathWithRange::new(citation);
65 let path = cx.agent_thread().update(cx, |thread, cx| {
66 thread
67 .project()
68 .read(cx)
69 .find_project_path(path_range.path.as_ref(), cx)
70 });
71
72 if let Ok(path) = cx.assert_some(path, format!("Valid path: {citation:?}"))
73 {
74 let buffer_text = {
75 let buffer = cx
76 .agent_thread()
77 .update(cx, |thread, cx| {
78 thread
79 .project()
80 .update(cx, |project, cx| project.open_buffer(path, cx))
81 })
82 .await
83 .ok();
84
85 let Ok(buffer_text) = cx.assert_some(
86 buffer.map(|buffer| {
87 buffer.read_with(cx, |buffer, _| buffer.text())
88 }),
89 "Reading buffer text succeeded",
90 ) else {
91 continue;
92 };
93 buffer_text
94 };
95
96 if let Some(content_len) = content_len {
97 // + 1 because there's a newline character after the citation.
98 let start_index = citation.len() + 1;
99 let end_index = content_len.saturating_sub(start_index);
100
101 if cx
102 .assert(
103 start_index <= end_index,
104 "Code block had a valid citation",
105 )
106 .is_ok()
107 {
108 let content = &text[start_index..end_index];
109
110 // deindent (trim the start of each line) because sometimes the model
111 // chooses to deindent its code snippets for the sake of readability,
112 // which in markdown is not only reasonable but usually desirable.
113 cx.assert(
114 deindent(&buffer_text)
115 .trim()
116 .contains(deindent(&content).trim()),
117 "Code block content was found in file",
118 )
119 .ok();
120
121 if let Some(range) = path_range.range {
122 let start_line_index = range.start.line.saturating_sub(1);
123 let line_count =
124 range.end.line.saturating_sub(start_line_index);
125 let mut snippet = buffer_text
126 .lines()
127 .skip(start_line_index as usize)
128 .take(line_count as usize)
129 .collect::<Vec<&str>>()
130 .join("\n");
131
132 if let Some(start_col) = range.start.col {
133 snippet = snippet[start_col as usize..].to_string();
134 }
135
136 if let Some(end_col) = range.end.col {
137 let last_line = snippet.lines().last().unwrap();
138 snippet = snippet[..snippet.len() - last_line.len()
139 + end_col as usize]
140 .to_string();
141 }
142
143 // deindent (trim the start of each line) because sometimes the model
144 // chooses to deindent its code snippets for the sake of readability,
145 // which in markdown is not only reasonable but usually desirable.
146 cx.assert_eq(
147 deindent(snippet.as_str()).trim(),
148 deindent(content).trim(),
149 format!(
150 "Code block was at {:?}-{:?}",
151 range.start, range.end
152 ),
153 )
154 .ok();
155 }
156 }
157 }
158 }
159 }
160 } else {
161 cx.assert(
162 false,
163 format!("Opening {FENCE} did not have a newline anywhere after it."),
164 )
165 .ok();
166 }
167
168 if let Some(content_len) = content_len {
169 // Advance past the closing backticks
170 text = &text[content_len + FENCE.len()..];
171 } else {
172 // There were no closing backticks associated with these opening backticks.
173 cx.assert(
174 false,
175 "Code block opening had matching closing backticks.".to_string(),
176 )
177 .ok();
178
179 // There are no more code blocks to parse, so we're done.
180 break;
181 }
182 }
183 }
184
185 Ok(())
186 }
187
188 fn thread_assertions(&self) -> Vec<JudgeAssertion> {
189 vec![
190 JudgeAssertion {
191 id: "trait method bodies are shown".to_string(),
192 description:
193 "All method bodies of the Tool trait are shown."
194 .to_string(),
195 },
196 JudgeAssertion {
197 id: "code blocks used".to_string(),
198 description:
199 "All code snippets are rendered inside markdown code blocks (as opposed to any other formatting besides code blocks)."
200 .to_string(),
201 },
202 JudgeAssertion {
203 id: "code blocks use backticks".to_string(),
204 description:
205 format!("All markdown code blocks use backtick fences ({FENCE}) rather than indentation.")
206 }
207 ]
208 }
209}
210
211fn deindent(as_str: impl AsRef<str>) -> String {
212 as_str
213 .as_ref()
214 .lines()
215 .map(|line| line.trim_start())
216 .collect::<Vec<&str>>()
217 .join("\n")
218}