refactor: 重构 AgentLoop 中的消息处理逻辑，简化工具结果处理和 LLM 请求构建
This commit is contained in:
oudecheng
parent
c8660df14b
commit
393052ae48
@ -917,7 +917,7 @@ impl AgentLoop {
|
||||
"Pre-process message state before sanitize"
|
||||
);
|
||||
}
|
||||
let removed = crate::bus::message::sanitize_incomplete_tool_call_sequences(&mut messages);
|
||||
let removed = Self::sanitize_messages_for_llm(&mut messages);
|
||||
if removed > 0 {
|
||||
tracing::warn!(
|
||||
removed_count = removed,
|
||||
@ -937,13 +937,11 @@ impl AgentLoop {
|
||||
// 检查取消信号
|
||||
// 使用 unwrap_or(true)：即使 watch channel 因异常关闭（sender drop 但未 send），
|
||||
// 也视为取消信号。defense-in-depth —— fail safe 而非 fail silent。
|
||||
if let Some(ref mutex) = self.cancel_token {
|
||||
if mutex.lock().await.has_changed().unwrap_or(true) {
|
||||
tracing::info!(iteration, "Agent execution cancelled by user");
|
||||
let cancel = Self::build_cancel_result(iteration, emitted_messages);
|
||||
self.emit_live_tool_call_message(cancel.final_response.clone()).await;
|
||||
return Ok(cancel);
|
||||
}
|
||||
if self.check_cancelled().await {
|
||||
tracing::info!(iteration, "Agent execution cancelled by user");
|
||||
let cancel = Self::build_cancel_result(iteration, emitted_messages);
|
||||
self.emit_live_tool_call_message(cancel.final_response.clone()).await;
|
||||
return Ok(cancel);
|
||||
}
|
||||
|
||||
// Build request
|
||||
@ -959,8 +957,7 @@ impl AgentLoop {
|
||||
// This catches edge cases where compression, persistence races,
|
||||
// or delta message merging may have introduced orphaned sequences
|
||||
// that survived the initial sanitization.
|
||||
let mid_loop_removed =
|
||||
crate::bus::message::sanitize_incomplete_tool_call_sequences(&mut messages);
|
||||
let mid_loop_removed = Self::sanitize_messages_for_llm(&mut messages);
|
||||
if mid_loop_removed > 0 {
|
||||
tracing::warn!(
|
||||
iteration = iteration,
|
||||
@ -969,50 +966,7 @@ impl AgentLoop {
|
||||
);
|
||||
}
|
||||
|
||||
// 过滤超出轮次和数量限制的图片
|
||||
let filtered_messages = filter_images_by_age_and_count(
|
||||
&messages,
|
||||
self.runtime_config.max_image_age_rounds,
|
||||
self.runtime_config.max_images_in_context,
|
||||
);
|
||||
let image_count = count_supported_image_media_refs(&filtered_messages);
|
||||
|
||||
// 构建系统提示词（统一注入 Agent 和 Skill 提示词）
|
||||
let system_prompt = system_prompt_context.and_then(|ctx| {
|
||||
self.system_prompt_provider
|
||||
.as_ref()
|
||||
.and_then(|provider| provider.build(ctx))
|
||||
});
|
||||
|
||||
let mut text_only_messages: Vec<Message> = Vec::with_capacity(filtered_messages.len() + 2);
|
||||
if let Some(ref prompt) = system_prompt {
|
||||
text_only_messages.push(Message::system(prompt.content.clone()));
|
||||
}
|
||||
text_only_messages.extend(filtered_messages.iter().map(chat_message_to_text_only_llm_message));
|
||||
|
||||
let image_tokens = image_token_budget_for_request(
|
||||
&self.runtime_config,
|
||||
&text_only_messages,
|
||||
tools.as_ref(),
|
||||
);
|
||||
let mut image_budget = ImageInlineBudget::new(image_tokens, image_count);
|
||||
let mut messages_for_llm: Vec<Message> = Vec::with_capacity(filtered_messages.len() + 2);
|
||||
// 使用相同的系统提示词（已构建）
|
||||
if let Some(ref prompt) = system_prompt {
|
||||
messages_for_llm.push(Message::system(prompt.content.clone()));
|
||||
}
|
||||
messages_for_llm.extend(
|
||||
filtered_messages
|
||||
.iter()
|
||||
.map(|message| chat_message_to_llm_message(message, &mut image_budget)),
|
||||
);
|
||||
|
||||
let request = ChatCompletionRequest {
|
||||
messages: messages_for_llm,
|
||||
temperature: None,
|
||||
max_tokens: None,
|
||||
tools,
|
||||
};
|
||||
let request = self.build_llm_request(&messages, system_prompt_context, tools);
|
||||
|
||||
// Set up streaming delta consumer
|
||||
// Pre-generate the message ID so stream deltas and the final assistant
|
||||
@ -1116,23 +1070,10 @@ impl AgentLoop {
|
||||
|
||||
// If no tool calls, this is the final response
|
||||
if response.tool_calls.is_empty() {
|
||||
let mut assistant_message = if let Some(reasoning_content) = response.reasoning_content
|
||||
{
|
||||
ChatMessage::assistant_with_reasoning(response.content, reasoning_content)
|
||||
} else {
|
||||
ChatMessage::assistant(response.content)
|
||||
};
|
||||
// Use the same ID as the stream deltas so the front-end can replace
|
||||
// the streamed message with this authoritative response.
|
||||
if had_streaming {
|
||||
assistant_message.id = streaming_message_id;
|
||||
}
|
||||
emitted_messages.push(assistant_message.clone());
|
||||
self.emit_live_tool_call_message(assistant_message.clone()).await;
|
||||
return Ok(AgentProcessResult {
|
||||
final_response: assistant_message,
|
||||
emitted_messages,
|
||||
});
|
||||
let result = self.build_final_response(
|
||||
response, &streaming_message_id, had_streaming, &mut emitted_messages,
|
||||
).await;
|
||||
return Ok(result);
|
||||
}
|
||||
|
||||
// Execute tool calls
|
||||
@ -1201,57 +1142,10 @@ impl AgentLoop {
|
||||
self.execute_tools(&response.tool_calls).await
|
||||
};
|
||||
|
||||
for (tool_call, result) in response.tool_calls.iter().zip(tool_results.iter()) {
|
||||
// Truncate tool result if too large
|
||||
let truncated_output =
|
||||
truncate_tool_result(&result.output, self.runtime_config.tool_result_max_chars);
|
||||
|
||||
// Record tool call and check for loops
|
||||
let loop_result = loop_detector.record(&tool_call.name, &tool_call.arguments);
|
||||
|
||||
match loop_result {
|
||||
LoopDetectionResult::Warning(msg) => {
|
||||
// Add warning and proceed
|
||||
tracing::warn!(
|
||||
tool = %tool_call.name,
|
||||
"Loop warning: {}",
|
||||
msg
|
||||
);
|
||||
let tool_message = ChatMessage::tool_with_state(
|
||||
tool_call.id.clone(),
|
||||
tool_call.name.clone(),
|
||||
format!("{}\n\n[上一条结果]\n{}", msg, truncated_output),
|
||||
if result.state == ToolExecutionState::PendingUserAction {
|
||||
ToolMessageState::PendingUserAction
|
||||
} else {
|
||||
ToolMessageState::Completed
|
||||
},
|
||||
)
|
||||
.with_tool_duration(result.duration.as_millis() as u64);
|
||||
messages.push(tool_message.clone());
|
||||
emitted_messages.push(tool_message.clone());
|
||||
let duration_ms = Some(result.duration.as_millis() as u64);
|
||||
self.emit_tool_result(tool_message, duration_ms).await;
|
||||
}
|
||||
LoopDetectionResult::Ok => {
|
||||
let tool_message = ChatMessage::tool_with_state(
|
||||
tool_call.id.clone(),
|
||||
tool_call.name.clone(),
|
||||
truncated_output,
|
||||
if result.state == ToolExecutionState::PendingUserAction {
|
||||
ToolMessageState::PendingUserAction
|
||||
} else {
|
||||
ToolMessageState::Completed
|
||||
},
|
||||
)
|
||||
.with_tool_duration(result.duration.as_millis() as u64);
|
||||
messages.push(tool_message.clone());
|
||||
emitted_messages.push(tool_message.clone());
|
||||
let duration_ms = Some(result.duration.as_millis() as u64);
|
||||
self.emit_tool_result(tool_message, duration_ms).await;
|
||||
}
|
||||
}
|
||||
}
|
||||
self.process_tool_results(
|
||||
&response.tool_calls, &tool_results, &mut loop_detector,
|
||||
&mut messages, &mut emitted_messages,
|
||||
).await;
|
||||
|
||||
// Loop continues to next iteration with updated messages
|
||||
// PendingUserAction 工具的结果已在上方加入 messages，
|
||||
@ -1264,116 +1158,8 @@ impl AgentLoop {
|
||||
);
|
||||
}
|
||||
|
||||
// Max iterations reached - ask LLM for a summary based on completed work
|
||||
tracing::warn!("Max iterations reached, requesting final summary from LLM");
|
||||
|
||||
// Defense: sanitize before final summary request
|
||||
let removed = crate::bus::message::sanitize_incomplete_tool_call_sequences(&mut messages);
|
||||
if removed > 0 {
|
||||
tracing::warn!(removed_count = removed, "Sanitized before max-iterations summary");
|
||||
}
|
||||
|
||||
// Add a message asking for summary
|
||||
let summary_request = ChatMessage::user(
|
||||
"You have reached the maximum number of tool call iterations. \
|
||||
Please provide your best answer based on the work completed so far.",
|
||||
);
|
||||
messages.push(summary_request);
|
||||
|
||||
// 过滤超出轮次和数量限制的图片
|
||||
let filtered_messages = filter_images_by_age_and_count(
|
||||
&messages,
|
||||
self.runtime_config.max_image_age_rounds,
|
||||
self.runtime_config.max_images_in_context,
|
||||
);
|
||||
|
||||
// Convert messages to LLM format (使用系统提示词提供者)
|
||||
let image_count = count_supported_image_media_refs(&filtered_messages);
|
||||
let mut text_only_messages: Vec<Message> = Vec::with_capacity(filtered_messages.len() + 1);
|
||||
if let Some(ref provider) = self.system_prompt_provider {
|
||||
if let Some(ctx) = system_prompt_context {
|
||||
if let Some(prompt) = provider.build(ctx) {
|
||||
text_only_messages.push(Message::system(prompt.content.clone()));
|
||||
}
|
||||
}
|
||||
}
|
||||
text_only_messages.extend(filtered_messages.iter().map(chat_message_to_text_only_llm_message));
|
||||
let image_tokens =
|
||||
image_token_budget_for_request(&self.runtime_config, &text_only_messages, None);
|
||||
let mut image_budget = ImageInlineBudget::new(image_tokens, image_count);
|
||||
let mut messages_for_llm: Vec<Message> = Vec::with_capacity(filtered_messages.len() + 1);
|
||||
if let Some(ref provider) = self.system_prompt_provider {
|
||||
if let Some(ctx) = system_prompt_context {
|
||||
if let Some(prompt) = provider.build(ctx) {
|
||||
messages_for_llm.push(Message::system(prompt.content.clone()));
|
||||
}
|
||||
}
|
||||
}
|
||||
messages_for_llm.extend(
|
||||
filtered_messages
|
||||
.iter()
|
||||
.map(|message| chat_message_to_llm_message(message, &mut image_budget)),
|
||||
);
|
||||
|
||||
let request = ChatCompletionRequest {
|
||||
messages: messages_for_llm,
|
||||
temperature: None,
|
||||
max_tokens: None,
|
||||
tools: None, // No tools in final summary call
|
||||
};
|
||||
|
||||
// 最终 summary 调用也与取消信号竞速
|
||||
let final_result: Result<
|
||||
crate::providers::ChatCompletionResponse,
|
||||
Box<dyn std::error::Error + Send + Sync>,
|
||||
>;
|
||||
if self.cancel_token.is_some() {
|
||||
tokio::select! {
|
||||
_ = self.cancel_signal() => {
|
||||
let cancel = Self::build_cancel_result(self.max_iterations, emitted_messages);
|
||||
self.emit_live_tool_call_message(cancel.final_response.clone()).await;
|
||||
return Ok(cancel);
|
||||
}
|
||||
result = self.provider.chat(request) => {
|
||||
final_result = result;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
final_result = self.provider.chat(request).await;
|
||||
}
|
||||
|
||||
match final_result {
|
||||
Ok(response) => {
|
||||
let assistant_message = if let Some(reasoning_content) = response.reasoning_content
|
||||
{
|
||||
ChatMessage::assistant_with_reasoning(response.content, reasoning_content)
|
||||
} else {
|
||||
ChatMessage::assistant(response.content)
|
||||
};
|
||||
emitted_messages.push(assistant_message.clone());
|
||||
self.emit_live_tool_call_message(assistant_message.clone()).await;
|
||||
Ok(AgentProcessResult {
|
||||
final_response: assistant_message,
|
||||
emitted_messages,
|
||||
})
|
||||
}
|
||||
Err(e) => {
|
||||
tracing::error!(
|
||||
provider = %self.provider.name(),
|
||||
model = %self.provider.model_id(),
|
||||
error = %e,
|
||||
error_details = %format_error_chain(e.as_ref()),
|
||||
"Failed to get summary from LLM"
|
||||
);
|
||||
let final_message = ChatMessage::assistant(recoverable_llm_message(&e.to_string()));
|
||||
emitted_messages.push(final_message.clone());
|
||||
self.emit_live_tool_call_message(final_message.clone()).await;
|
||||
Ok(AgentProcessResult {
|
||||
final_response: final_message,
|
||||
emitted_messages,
|
||||
})
|
||||
}
|
||||
}
|
||||
// Max iterations reached - request final summary from LLM
|
||||
Ok(self.run_final_summary(&mut messages, system_prompt_context, &mut emitted_messages).await)
|
||||
}
|
||||
|
||||
/// 等待取消信号。若未配置 cancel_token，永远不返回。
|
||||
@ -1389,6 +1175,217 @@ impl AgentLoop {
|
||||
}
|
||||
}
|
||||
|
||||
/// 净化消息中的不完整 tool_call 序列，返回移除数量。
|
||||
/// 日志由调用方负责，因为不同调用点需要不同的结构化字段。
|
||||
fn sanitize_messages_for_llm(messages: &mut Vec<ChatMessage>) -> usize {
|
||||
crate::bus::message::sanitize_incomplete_tool_call_sequences(messages)
|
||||
}
|
||||
|
||||
/// 检查取消信号。返回 true 表示已取消。
|
||||
async fn check_cancelled(&self) -> bool {
|
||||
if let Some(ref mutex) = self.cancel_token {
|
||||
mutex.lock().await.has_changed().unwrap_or(true)
|
||||
} else {
|
||||
false
|
||||
}
|
||||
}
|
||||
|
||||
/// 构建 LLM 请求：过滤图片、构建系统提示、转换消息格式。
|
||||
/// 统一了循环内调用和最终 summary 调用的请求构建逻辑。
|
||||
fn build_llm_request(
|
||||
&self,
|
||||
messages: &[ChatMessage],
|
||||
system_prompt_context: Option<&SystemPromptContext>,
|
||||
tools: Option<Vec<crate::domain::tools::Tool>>,
|
||||
) -> ChatCompletionRequest {
|
||||
let filtered_messages = filter_images_by_age_and_count(
|
||||
messages,
|
||||
self.runtime_config.max_image_age_rounds,
|
||||
self.runtime_config.max_images_in_context,
|
||||
);
|
||||
let image_count = count_supported_image_media_refs(&filtered_messages);
|
||||
|
||||
let system_prompt = system_prompt_context.and_then(|ctx| {
|
||||
self.system_prompt_provider
|
||||
.as_ref()
|
||||
.and_then(|provider| provider.build(ctx))
|
||||
});
|
||||
|
||||
let mut text_only_messages: Vec<Message> = Vec::with_capacity(filtered_messages.len() + 2);
|
||||
if let Some(ref prompt) = system_prompt {
|
||||
text_only_messages.push(Message::system(prompt.content.clone()));
|
||||
}
|
||||
text_only_messages.extend(filtered_messages.iter().map(chat_message_to_text_only_llm_message));
|
||||
|
||||
let image_tokens = image_token_budget_for_request(
|
||||
&self.runtime_config,
|
||||
&text_only_messages,
|
||||
tools.as_ref(),
|
||||
);
|
||||
let mut image_budget = ImageInlineBudget::new(image_tokens, image_count);
|
||||
let mut messages_for_llm: Vec<Message> = Vec::with_capacity(filtered_messages.len() + 2);
|
||||
if let Some(ref prompt) = system_prompt {
|
||||
messages_for_llm.push(Message::system(prompt.content.clone()));
|
||||
}
|
||||
messages_for_llm.extend(
|
||||
filtered_messages
|
||||
.iter()
|
||||
.map(|message| chat_message_to_llm_message(message, &mut image_budget)),
|
||||
);
|
||||
|
||||
ChatCompletionRequest {
|
||||
messages: messages_for_llm,
|
||||
temperature: None,
|
||||
max_tokens: None,
|
||||
tools,
|
||||
}
|
||||
}
|
||||
|
||||
/// 构建 LLM 最终响应（无工具调用时）。调用方应先检查 `response.tool_calls.is_empty()`。
|
||||
async fn build_final_response(
|
||||
&self,
|
||||
response: crate::providers::ChatCompletionResponse,
|
||||
streaming_message_id: &str,
|
||||
had_streaming: bool,
|
||||
emitted_messages: &mut Vec<ChatMessage>,
|
||||
) -> AgentProcessResult {
|
||||
let mut assistant_message = if let Some(reasoning_content) = response.reasoning_content {
|
||||
ChatMessage::assistant_with_reasoning(response.content, reasoning_content)
|
||||
} else {
|
||||
ChatMessage::assistant(response.content)
|
||||
};
|
||||
// Use the same ID as the stream deltas so the front-end can replace
|
||||
// the streamed message with this authoritative response.
|
||||
if had_streaming {
|
||||
assistant_message.id = streaming_message_id.to_string();
|
||||
}
|
||||
emitted_messages.push(assistant_message.clone());
|
||||
self.emit_live_tool_call_message(assistant_message.clone()).await;
|
||||
AgentProcessResult {
|
||||
final_response: assistant_message,
|
||||
emitted_messages: std::mem::take(emitted_messages),
|
||||
}
|
||||
}
|
||||
|
||||
/// 处理工具执行结果，追加到 messages 和 emitted_messages，发出 tool_result 事件。
|
||||
/// 合并了 LoopDetectionResult::Warning 和 ::Ok 两个分支的逻辑。
|
||||
async fn process_tool_results(
|
||||
&self,
|
||||
tool_calls: &[ToolCall],
|
||||
tool_results: &[ToolExecutionOutcome],
|
||||
loop_detector: &mut LoopDetector,
|
||||
messages: &mut Vec<ChatMessage>,
|
||||
emitted_messages: &mut Vec<ChatMessage>,
|
||||
) {
|
||||
for (tool_call, result) in tool_calls.iter().zip(tool_results.iter()) {
|
||||
let truncated_output =
|
||||
truncate_tool_result(&result.output, self.runtime_config.tool_result_max_chars);
|
||||
|
||||
let loop_result = loop_detector.record(&tool_call.name, &tool_call.arguments);
|
||||
let prefix = match loop_result {
|
||||
LoopDetectionResult::Warning(msg) => {
|
||||
tracing::warn!(tool = %tool_call.name, "Loop warning: {}", msg);
|
||||
format!("{}\n\n[上一条结果]\n", msg)
|
||||
}
|
||||
LoopDetectionResult::Ok => String::new(),
|
||||
};
|
||||
|
||||
let state = if result.state == ToolExecutionState::PendingUserAction {
|
||||
ToolMessageState::PendingUserAction
|
||||
} else {
|
||||
ToolMessageState::Completed
|
||||
};
|
||||
let tool_message = ChatMessage::tool_with_state(
|
||||
tool_call.id.clone(),
|
||||
tool_call.name.clone(),
|
||||
format!("{}{}", prefix, truncated_output),
|
||||
state,
|
||||
)
|
||||
.with_tool_duration(result.duration.as_millis() as u64);
|
||||
messages.push(tool_message.clone());
|
||||
emitted_messages.push(tool_message.clone());
|
||||
let duration_ms = Some(result.duration.as_millis() as u64);
|
||||
self.emit_tool_result(tool_message, duration_ms).await;
|
||||
}
|
||||
}
|
||||
|
||||
/// 达到最大迭代次数后，请求 LLM 生成总结。
|
||||
async fn run_final_summary(
|
||||
&self,
|
||||
messages: &mut Vec<ChatMessage>,
|
||||
system_prompt_context: Option<&SystemPromptContext>,
|
||||
emitted_messages: &mut Vec<ChatMessage>,
|
||||
) -> AgentProcessResult {
|
||||
tracing::warn!("Max iterations reached, requesting final summary from LLM");
|
||||
|
||||
// Defense: sanitize before final summary request
|
||||
let removed = Self::sanitize_messages_for_llm(messages);
|
||||
if removed > 0 {
|
||||
tracing::warn!(removed_count = removed, "Sanitized before max-iterations summary");
|
||||
}
|
||||
|
||||
// Add a message asking for summary
|
||||
let summary_request = ChatMessage::user(
|
||||
"You have reached the maximum number of tool call iterations. \
|
||||
Please provide your best answer based on the work completed so far.",
|
||||
);
|
||||
messages.push(summary_request);
|
||||
|
||||
let request = self.build_llm_request(messages, system_prompt_context, None);
|
||||
|
||||
// 最终 summary 调用也与取消信号竞速
|
||||
let final_result: Result<
|
||||
crate::providers::ChatCompletionResponse,
|
||||
Box<dyn std::error::Error + Send + Sync>,
|
||||
>;
|
||||
if self.cancel_token.is_some() {
|
||||
tokio::select! {
|
||||
_ = self.cancel_signal() => {
|
||||
let cancel = Self::build_cancel_result(self.max_iterations, std::mem::take(emitted_messages));
|
||||
self.emit_live_tool_call_message(cancel.final_response.clone()).await;
|
||||
return cancel;
|
||||
}
|
||||
result = self.provider.chat(request) => {
|
||||
final_result = result;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
final_result = self.provider.chat(request).await;
|
||||
}
|
||||
|
||||
match final_result {
|
||||
Ok(response) => {
|
||||
let assistant_message = if let Some(reasoning_content) = response.reasoning_content {
|
||||
ChatMessage::assistant_with_reasoning(response.content, reasoning_content)
|
||||
} else {
|
||||
ChatMessage::assistant(response.content)
|
||||
};
|
||||
emitted_messages.push(assistant_message.clone());
|
||||
self.emit_live_tool_call_message(assistant_message.clone()).await;
|
||||
AgentProcessResult {
|
||||
final_response: assistant_message,
|
||||
emitted_messages: std::mem::take(emitted_messages),
|
||||
}
|
||||
}
|
||||
Err(e) => {
|
||||
tracing::error!(
|
||||
provider = %self.provider.name(),
|
||||
model = %self.provider.model_id(),
|
||||
error = %e,
|
||||
error_details = %format_error_chain(e.as_ref()),
|
||||
"Failed to get summary from LLM"
|
||||
);
|
||||
let final_message = ChatMessage::assistant(recoverable_llm_message(&e.to_string()));
|
||||
emitted_messages.push(final_message.clone());
|
||||
self.emit_live_tool_call_message(final_message.clone()).await;
|
||||
AgentProcessResult {
|
||||
final_response: final_message,
|
||||
emitted_messages: std::mem::take(emitted_messages),
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// 构建取消响应，包含已完成的迭代次数和已生成的消息数量。
|
||||
/// 构建取消响应，将取消通知追加到 `emitted_messages` 末尾后一并返回。
|
||||
/// 这样 finalize_result 会把中间消息加入内存历史，确保下一个 LLM 调用有完整上下文。
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user