//! Compact Prompt 模板
//!
//! L1/L2/L3 各层的 LLM prompt 设计
use super::source::{Talk, ToolCall};
/// L1: Observation 生成 Prompt
pub struct L1Prompt;
impl L1Prompt {
/// 系统提示
pub const SYSTEM: &'static str = r#"You are a code operation analyst. Analyze tool calls and generate structured observations.
Output format (JSON):
{
"type": "bugfix|feature|refactor|change|discovery|decision",
"title": "Short title (< 20 words)",
"subtitle": "One-line explanation (optional)",
"facts": ["Specific fact 1", "Specific fact 2"],
"narrative": "Full context description (optional, for complex operations)"
}
Type definitions:
- bugfix: Bug fix
- feature: New feature
- refactor: Code refactoring
- change: General modification
- discovery: Code discovery/understanding
- decision: Technical decision
Requirements:
- Keep it concise, capture the core
- facts should contain searchable keywords
- narrative can be omitted for simple operations
- Output language: Match the language of the input content"#;
/// 构建用户消息
pub fn build_user_message(tool_call: &ToolCall, context: Option<&str>) -> String {
let mut prompt = String::new();
prompt.push_str(&format!("Tool: {}\n", tool_call.name));
if let Some(args) = &tool_call.arguments {
// 截断过长的参数
let args_display = if args.len() > 1000 {
format!("{}...(truncated)", &args[..1000])
} else {
args.clone()
};
prompt.push_str(&format!("Arguments: {}\n", args_display));
}
if let Some(output) = &tool_call.output {
// 截断过长的输出
let output_display = if output.len() > 2000 {
format!("{}...(truncated)", &output[..2000])
} else {
output.clone()
};
prompt.push_str(&format!("Output: {}\n", output_display));
}
if let Some(ctx) = context {
prompt.push_str(&format!("\nContext: {}\n", ctx));
}
prompt
}
}
/// L2: Talk Summary 生成 Prompt
pub struct L2Prompt;
impl L2Prompt {
/// 系统提示
pub const SYSTEM: &'static str = r#"You are a conversation summarizer. Compress one round of AI coding conversation into a concise summary.
Output format (JSON):
{
"user_request": "Brief description of user request (< 50 words)",
"summary": "What was accomplished in this round (< 200 words)",
"completed": "Specific completed items (optional)",
"files_involved": ["File paths involved"]
}
Requirements:
- Preserve key decisions and solutions
- Remove duplicates and redundancy
- Maintain technical accuracy
- Extract files_involved from tool calls
- Output language: Match the language of the input content"#;
/// 构建用户消息
pub fn build_user_message(talk: &Talk) -> String {
let mut prompt = String::new();
// 用户请求
prompt.push_str(&format!(
"[User]: {}\n\n",
truncate(&talk.user_message.content, 2000)
));
// Assistant 回复
for msg in &talk.assistant_messages {
prompt.push_str(&format!(
"[Assistant]: {}\n\n",
truncate(&msg.content, 2000)
));
}
// 工具调用摘要
if !talk.tool_calls.is_empty() {
prompt.push_str("Tool calls:\n");
for tc in &talk.tool_calls {
prompt.push_str(&format!("- {}", tc.name));
if let Some(path) = tc.extract_file_path() {
prompt.push_str(&format!(" ({})", path));
}
prompt.push('\n');
}
}
prompt
}
}
/// L3: Session Summary 生成 Prompt
pub struct L3Prompt;
impl L3Prompt {
/// 系统提示
pub const SYSTEM: &'static str = r#"You are a session summarizer. Consolidate multiple conversation round summaries into a complete session summary.
Output format (JSON):
{
"summary": "One-sentence summary of the entire session (< 100 words)",
"key_points": ["Key point 1", "Key point 2", ...],
"files_involved": ["Main files involved"],
"technologies": ["Technologies/tools involved"]
}
Requirements:
- summary should quickly convey what this session accomplished
- key_points: 3-7 items covering main work
- Remove duplicates, merge related content
- Maintain technical accuracy
- Output language: Match the language of the input content"#;
/// 从 L2 Talk Summaries 构建用户消息
pub fn build_user_message(
talk_summaries: &[(i32, String, Option<String>)], // (prompt_number, summary, completed)
project_path: Option<&str>,
) -> String {
let mut prompt = String::new();
if let Some(path) = project_path {
prompt.push_str(&format!("Project: {}\n\n", path));
}
for (_prompt_number, summary, completed) in talk_summaries {
prompt.push_str(&format!("- {}", summary));
if let Some(comp) = completed {
prompt.push_str(&format!(". {}", comp));
}
prompt.push('\n');
}
prompt
}
}
/// 截断字符串
fn truncate(s: &str, max_len: usize) -> &str {
if s.len() <= max_len {
s
} else {
// 找到 UTF-8 安全的截断点
let mut end = max_len;
while end > 0 && !s.is_char_boundary(end) {
end -= 1;
}
&s[..end]
}
}
/// L1 Observation 输出结构
#[derive(Debug, Clone, serde::Deserialize)]
pub struct L1Output {
#[serde(rename = "type")]
pub observation_type: String,
pub title: String,
pub subtitle: Option<String>,
pub facts: Option<Vec<String>>,
pub narrative: Option<String>,
}
/// L2 Talk Summary 输出结构
#[derive(Debug, Clone, serde::Deserialize)]
pub struct L2Output {
pub user_request: Option<String>,
pub summary: String,
/// 完成事项（LLM 可能返回 string 或 array，用 Value 兼容）
#[serde(default)]
pub completed: Option<serde_json::Value>,
pub files_involved: Option<Vec<String>>,
}
impl L2Output {
/// 获取 completed 字符串（兼容 string/array 格式）
pub fn completed_str(&self) -> Option<String> {
match &self.completed {
Some(serde_json::Value::String(s)) if !s.is_empty() => Some(s.clone()),
Some(serde_json::Value::Array(arr)) if !arr.is_empty() => {
let items: Vec<String> = arr
.iter()
.filter_map(|v| v.as_str().map(|s| s.to_string()))
.collect();
if items.is_empty() {
None
} else {
Some(items.join("; "))
}
}
_ => None,
}
}
}
/// L3 Session Summary 输出结构
#[derive(Debug, Clone, serde::Deserialize)]
pub struct L3Output {
pub summary: String,
pub key_points: Option<Vec<String>>,
pub files_involved: Option<Vec<String>>,
pub technologies: Option<Vec<String>>,
}
#[cfg(test)]
mod tests {
use super::*;
use crate::compact::source::{Message, MessageRole};
#[test]
fn test_truncate() {
assert_eq!(truncate("hello", 10), "hello");
assert_eq!(truncate("hello world", 5), "hello");
// UTF-8 安全
assert_eq!(truncate("你好世界", 6), "你好"); // 每个中文 3 bytes
}
#[test]
fn test_l1_prompt_build() {
let tc = ToolCall {
id: "1".to_string(),
name: "Read".to_string(),
arguments: Some(r#"{"file_path": "/src/main.rs"}"#.to_string()),
output: Some("fn main() { }".to_string()),
timestamp: 0,
sequence: 0,
prompt_number: 1,
};
let prompt = L1Prompt::build_user_message(&tc, None);
assert!(prompt.contains("Read"));
assert!(prompt.contains("file_path"));
}
#[test]
fn test_l2_prompt_build() {
let talk = Talk {
prompt_number: 1,
user_message: Message {
id: 1,
uuid: "1".to_string(),
role: MessageRole::User,
content: "Help me fix the bug".to_string(),
timestamp: 0,
sequence: 0,
prompt_number: 1,
tool_name: None,
tool_args: None,
},
assistant_messages: vec![Message {
id: 2,
uuid: "2".to_string(),
role: MessageRole::Assistant,
content: "I found the issue...".to_string(),
timestamp: 0,
sequence: 1,
prompt_number: 1,
tool_name: None,
tool_args: None,
}],
tool_calls: vec![],
};
let prompt = L2Prompt::build_user_message(&talk);
assert!(prompt.contains("[User]: Help me fix the bug"));
assert!(prompt.contains("[Assistant]: I found the issue"));
}
#[test]
fn test_l3_prompt_build() {
let summaries = vec![
(
1,
"Fixed authentication bug".to_string(),
Some("Added null check".to_string()),
),
(2, "Added new API endpoint".to_string(), None),
];
let prompt = L3Prompt::build_user_message(&summaries, Some("/project/path"));
assert!(prompt.contains("Project: /project/path"));
assert!(prompt.contains("- Fixed authentication bug. Added null check"));
assert!(prompt.contains("- Added new API endpoint"));
}
}
