lxy/DevFlow
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

新增: F-01阶段5模型路由调用点接入

Browse Source
This commit is contained in:
绝尘
2026-06-17 00:33:44 +08:00
parent 5b7f342e14
commit 2ddfea67c7
10 changed files with 220 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
src-tauri/src/commands/ai/agentic.rs
View File

@@ -12,6 +12,13 @@ use df_ai::provider::{ChatMessage, CompletionRequest, LlmProvider};
// CR-30-1: 复用 retry::backoff_delay(jitter 1s→2s→4s) + is_status_retryable(Fatal 分类)
// 实现流前失败重试退避对齐(决策 F-260616-07 a1),避免重写退避逻辑。
use df_ai::retry;
// F-01 阶段5: 智能路由 helper + TaskRequirements + 维度枚举。
// 调用点构造 TaskRequirements(主对话:needs_tool_use=true,min_intelligence=Standard,
// 当前仅 Text 模态;后续多模态接入时检测消息内 Part/Image 追加 Vision),
// 经 select_model_id 在 provider.model_configs 池中选最优;池空/无匹配兜底 default_model。
use df_ai::router::{
select_model_id, IntelligenceTier, Modality, TaskRequirements,
};
use df_storage::db::Database;
use df_storage::models::AiProviderRecord;
@@ -154,6 +161,19 @@ pub(crate) async fn run_agentic_loop(
key_len = key_len,
"[ai] 发起 LLM 请求"
);
// F-01 阶段5: 主对话路由 — TaskRequirements(needs_tool_use=true,min_intelligence=Standard)。
// 模态当前仅 Text(图像消息类型未实现,后续多模态接入时检测 Part/Image 追加 Vision)。
// select_model_id None(池空/无匹配)→ 兜底 default_model,行为与接入前一致。
let agentic_req = TaskRequirements {
modalities: vec![Modality::Text],
needs_tool_use: true,
min_intelligence: IntelligenceTier::Standard,
max_cost: None,
estimated_context: 0,
};
let resolved_model = select_model_id(&agentic_req, &provider_config.model_configs)
.unwrap_or_else(|| provider_config.default_model.clone());
let tool_defs = tools_arc.tool_definitions();
// 停止信号副本：stream_llm 与每轮迭代共享读取，避免重复加锁
// notify 同取一份 Arc 引用（B-260615-14）：stream_llm select! 监听 notified() 即时唤醒
@@ -265,7 +285,7 @@ pub(crate) async fn run_agentic_loop(
// 一致,重建与复用等价。收益:省去每次重试的 lock() + build_for_request(含
// all_messages_clone 全量 clone)。
let retry_request = CompletionRequest {
model: provider_config.default_model.clone(),
model: resolved_model.clone(),
messages: messages.clone(),
temperature: Some(0.7),
max_tokens: Some(8192),
@@ -387,16 +407,16 @@ pub(crate) async fn run_agentic_loop(
}
if !full_text.is_empty() {
let mut msg = ChatMessage::assistant(&full_text);
msg.model = Some(provider_config.default_model.clone());
msg.model = Some(resolved_model.clone());
session.messages.push(msg);
// 追加系统提示消息:响应因网络中断不完整(对齐决策 a1 系统提示机制)
let mut notice = ChatMessage::system("⚠ 响应因网络中断不完整,以上为已接收的部分内容。可重新发送以获取完整回复。");
notice.model = Some(provider_config.default_model.clone());
notice.model = Some(resolved_model.clone());
session.messages.push(notice);
}
}
save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&provider_config.default_model)).await;
save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&resolved_model)).await;
// 标题生成后台化(失败有 extract_title 兜底)
spawn_ensure_title(&provider_config, &db, &conv_id, &app_handle, &session_arc, &llm_concurrency);
guard.reset().await;
@@ -442,11 +462,11 @@ pub(crate) async fn run_agentic_loop(
})
.collect();
let mut msg = ChatMessage::assistant_with_tools(&full_text, ai_tool_calls);
msg.model = Some(provider_config.default_model.clone());
msg.model = Some(resolved_model.clone());
session.messages.push(msg);
} else if !full_text.is_empty() {
let mut msg = ChatMessage::assistant(&full_text);
msg.model = Some(provider_config.default_model.clone());
msg.model = Some(resolved_model.clone());
session.messages.push(msg);
}
}
@@ -458,7 +478,7 @@ pub(crate) async fn run_agentic_loop(
completion_tokens: tokens.completion(),
total_tokens: tokens.total(),
};
save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&provider_config.default_model)).await;
save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&resolved_model)).await;
// 标题生成后台化：不阻塞 Completed emit（失败有 extract_title 兜底）
spawn_ensure_title(&provider_config, &db, &conv_id, &app_handle, &session_arc, &llm_concurrency);
guard.reset().await;
@@ -483,7 +503,7 @@ pub(crate) async fn run_agentic_loop(
completion_tokens: tokens.completion(),
total_tokens: tokens.total(),
};
save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&provider_config.default_model)).await;
save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&resolved_model)).await;
// B-260615-26: 审批等待 return 前 disarm guard——保持 generating=true 留 try_continue 续生成,
// 同时 Drop 因 done=true 跳过复位 spawn(避免误复位审批态 generating 致 ai_approve→try_continue 不续)
guard.disarm();
@@ -512,7 +532,7 @@ pub(crate) async fn run_agentic_loop(
completion_tokens: tokens.completion(),
total_tokens: tokens.total(),
};
save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&provider_config.default_model)).await;
save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&resolved_model)).await;
// 暂停态保持 generating=true(防其他 send 抢占,仿审批),disarm guard 跳过 Drop 兜底复位
guard.disarm();
let _ = app_handle.emit("ai-chat-event", AiChatEvent::AiMaxRoundsReached {
@@ -540,8 +560,9 @@ pub(crate) async fn run_agentic_loop(
let knowledge_config = knowledge_config.clone();
let app_handle = app_handle.clone();
let llm_concurrency = llm_concurrency.clone();
let resolved_model = resolved_model.clone();
tauri::async_runtime::spawn(async move {
save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&provider_config.default_model)).await;
save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&resolved_model)).await;
// 知识提炼:需读已落库的对话消息,故在 save 之后
if let Err(e) = maybe_spawn_extraction(&session_arc, &db, &conv_id, &provider_config, &knowledge_config, llm_concurrency.clone()).await {
tracing::warn!("知识提炼触发失败(非阻断): {}", e);