From 08e9d0292d15792666eca79e6c17db932879a28c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E7=BB=9D=E5=B0=98?= <237809796@qq.com> Date: Sun, 4 Oct 2026 16:06:08 +0800 Subject: [PATCH] =?UTF-8?q?=E4=BC=98=E5=8C=96:=20=E5=AD=90=E4=BB=A3?= =?UTF-8?q?=E7=90=86=E6=88=AA=E6=96=AD=E4=BF=A1=E5=8F=B7=E4=B8=8E=E7=BC=93?= =?UTF-8?q?=E5=AD=98=E5=A4=B1=E6=95=88?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - max_tokens 截断标记穿透到完成卡 - 缓存写后失效补测试锁定 - 派发参数卡头信息补全 --- .../src/commands/ai/agentic/stream_relay.rs | 37 ++++++--- src-tauri/src/commands/ai/audit/cache.rs | 82 ++++++++++++++++++- .../src/commands/ai/augmentation/registry.rs | 55 ++++++++++++- src/composables/ai/useAiSend.ts | 4 +- src/i18n/en/aiTool.ts | 2 + src/i18n/zh-CN/aiTool.ts | 2 + 6 files changed, 165 insertions(+), 17 deletions(-) diff --git a/src-tauri/src/commands/ai/agentic/stream_relay.rs b/src-tauri/src/commands/ai/agentic/stream_relay.rs index b021a21..335a776 100644 --- a/src-tauri/src/commands/ai/agentic/stream_relay.rs +++ b/src-tauri/src/commands/ai/agentic/stream_relay.rs @@ -69,6 +69,9 @@ enum StreamOutcome { resolved_model: String, /// DeepSeek thinking 模式推理内容（需回传到下一轮请求） reasoning_content: Option, + /// 流因 provider max_tokens 上限截断(finish_reason=length / stop_reason=max_tokens), + /// 本轮文本/tool_calls 可能不完整,调用方据此标记该轮产出截断。 + truncated: bool, }, /// error 字段供外层全 candidate 耗尽时 emit 最终 AiError(单气泡聚合)。 InitFailedExhausted { error: String }, @@ -175,10 +178,11 @@ async fn stream_with_retry( ).await; match result { - StreamResult::Complete { text, tool_calls, usage, reasoning_content } => { + StreamResult::Complete { text, tool_calls, usage, reasoning_content, truncated_by_max_tokens } => { return StreamOutcome::Success { text, tool_calls, usage, incomplete: false, resolved_model: resolved_model.to_string(), reasoning_content, + truncated: truncated_by_max_tokens, }; } StreamResult::Partial { text, tool_calls, usage, reasoning_content } => { @@ -192,6 +196,8 @@ async fn stream_with_retry( return StreamOutcome::Success { text, tool_calls, usage, incomplete: true, resolved_model: resolved_model.to_string(), reasoning_content, + // MidStream 已按不完整处理(incomplete=true),截断标志不再叠加 + truncated: false, }; } StreamResult::InitFailed { retryable, error } => { @@ -257,6 +263,9 @@ pub(crate) enum RoundOutcome { /// 本轮 prompt 是否估算兜底(= 进 stream_round 时的 round_usage.prompt_tokens == 0), /// 供 loop 主体 push 消息级 is_estimated,与 round_usage 同点派生防漂移。 round_is_estimated: bool, + /// 本轮流因 provider max_tokens 上限截断(P2-1),文本/tool_calls 可能不完整。 + /// 子循环据此把收敛报告标 Truncated,不冒充 Completed;主循环据此提示用户。 + truncated_by_max_tokens: bool, }, /// 流返回后 stale(epoch 已变,被新 loop 接管):guard 已 disarm, /// loop 应直接 return(不再 push/save/执行工具)。 @@ -265,8 +274,9 @@ pub(crate) enum RoundOutcome { /// teardown_round save/reset/emit)→ loop 直接 return。 MidStreamSaved, /// Fatal(候选 Fatal 或全候选耗尽):guard.reset + epoch 检查 + emit AiError + - /// save 已完成 → loop 直接 return。 - Fatal, + /// save 已完成 → loop 直接 return。`error` 为最后一条诊断文本(P2-4:子循环据此 + /// 在出口报告写明失败原因,父会话可针对性补救,不再拿到无因失败)。 + Fatal { error: String }, } /// 单轮流式调用:抽自 `run_agentic_loop` for 循环体的「单轮流式」自包含块。 @@ -365,9 +375,9 @@ pub(crate) async fn stream_round( ); } - let (full_text, tool_calls_acc, round_usage, incomplete, round_reasoning_content) = { + let (full_text, tool_calls_acc, round_usage, incomplete, round_reasoning_content, round_truncated) = { // outcome 累积成功结果(Complete/Partial),InitFailed 不写入。 - let mut outcome: Option<(String, std::collections::HashMap, df_ai::provider::TokenUsage, bool, Option)> = None; + let mut outcome: Option<(String, std::collections::HashMap, df_ai::provider::TokenUsage, bool, Option, bool)> = None; // 追踪最后一个 Exhausted candidate 的诊断文本,供「全 candidate 耗尽」时 emit 最终 AiError。 // Fatal 分支即时 emit(终态不切候选),Exhausted 分支仅记录 error 不 emit(可能切下一 candidate 成功, // emit 会留残留气泡)。 @@ -418,14 +428,14 @@ pub(crate) async fn stream_round( ).await; match stream_outcome { - StreamOutcome::Success { text, tool_calls, usage, incomplete, resolved_model: m, reasoning_content: rc } => { + StreamOutcome::Success { text, tool_calls, usage, incomplete, resolved_model: m, reasoning_content: rc, truncated } => { // 成功:更新迭代级 resolved_model + provider_config(供后续 push/save/标题)。 // mut 解构(已在顶部声明 mut):本轮后续 save/push 用「实际成功所用 provider」 // 而非主 candidate。compress 已在本轮 stream 之前用过 primary,不受影响。 *resolved_model = m; *provider_config = candidate.clone(); *last_reasoning_content = rc.clone(); - outcome = Some((text, tool_calls, usage, incomplete, last_reasoning_content.clone())); + outcome = Some((text, tool_calls, usage, incomplete, last_reasoning_content.clone(), truncated)); break 'candidate; } StreamOutcome::InitFailedExhausted { error } => { @@ -449,14 +459,14 @@ pub(crate) async fn stream_round( guard.reset().await; // F1:旧 loop(被新 loop 接管)不 emit 错误气泡(owner loop 负责呈现)。 if loop_epoch_arc.load(Ordering::SeqCst) != my_epoch { - return RoundOutcome::Fatal; + return RoundOutcome::Fatal { error }; } emit_fatal_error(app_handle, conv_id, &error).await; // G1.1(2026-08-05):Fatal 退出也落库 user 消息(镜像下方 Exhausted 分支), // 否则重启/切会话从 DB 恢复时末条 user 丢失(与 Exhausted 同根因不对称)。 // epoch 校验通过后、return 前补 save(owner loop 才写,防 owner 已转移仍写)。 save_conversation(session_arc, db, conv_id, None, Some(&*resolved_model), true).await; - return RoundOutcome::Fatal; + return RoundOutcome::Fatal { error }; } } } @@ -471,16 +481,16 @@ pub(crate) async fn stream_round( "[ai] 全 provider 候选流前失败重试耗尽,放弃本轮", ); guard.reset().await; - if loop_epoch_arc.load(Ordering::SeqCst) != my_epoch { - return RoundOutcome::Fatal; - } let err_msg = last_exhausted_error .unwrap_or_else(|| "AI 调用失败：所有候选 provider 重试耗尽".to_string()); + if loop_epoch_arc.load(Ordering::SeqCst) != my_epoch { + return RoundOutcome::Fatal { error: err_msg }; + } emit_fatal_error(app_handle, conv_id, &err_msg).await; // 全 provider 失败也需落库 user 消息(已 push 内存), // 否则切换/重载从 DB 恢复时末条 user 丢失(实测压缩触发失败场景:DB 末条 tool 无后续 user)。 save_conversation(session_arc, db, conv_id, None, Some(&*resolved_model), true).await; - return RoundOutcome::Fatal; + return RoundOutcome::Fatal { error: err_msg }; } } }; @@ -584,6 +594,7 @@ pub(crate) async fn stream_round( tool_calls: tool_calls_acc, round_usage: round_usage_full, round_is_estimated, + truncated_by_max_tokens: round_truncated, } } diff --git a/src-tauri/src/commands/ai/audit/cache.rs b/src-tauri/src/commands/ai/audit/cache.rs index 3f51bb9..4a6bb00 100644 --- a/src-tauri/src/commands/ai/audit/cache.rs +++ b/src-tauri/src/commands/ai/audit/cache.rs @@ -257,6 +257,11 @@ pub(crate) fn readonly_cache_args_key(tool_name: &str, args: &serde_json::Value) /// /// 与 find_cached_high_risk_result 区别:后者治"超时→重试→重新审批"循环( High risk), /// 本函数治"LLM 失忆→同参重调"死循环( Low risk 只读),问题形态相似但触发原因不同。 +/// +/// **写后失效**(治「只读结果缓存无写失效」缺陷):会话内 write_file/run_command 等 +/// 非只读工具成功执行后,历史只读结果已可能过期(读的是改前世界)。本函数扫描时按 +/// per_conv.readonly_cache_floor 过滤——floor 之后的索引才算可信,之前的旧调用全部 +/// 跳过(过界即脏),写后 LLM 重读同参工具必然 miss 缓存走真执行拿到新内容。 pub(crate) async fn find_cached_readonly_result( session_arc: &Arc>, conv_id: &str, @@ -281,7 +286,11 @@ pub(crate) async fn find_cached_readonly_result( let cached: Option<(String, String, u32)> = (async { let session = session_arc.lock().await; let Some(conv) = session.conv_read(conv_id) else { return None }; - let msgs: Vec<&ChatMessage> = conv.messages.iter().collect(); + // 写后失效:floor 之前的只读结果一律不可信(写过世界状态),只扫 [floor, len)。 + // floor > len 防御(pop_last_assistant_round/clear 缩短历史后索引漂移),钳到 len + // 等于扫描区间空 → 无缓存可命中,走真执行(保守方向安全)。 + let floor = conv.readonly_cache_floor.min(conv.messages.len()); + let msgs: Vec<&ChatMessage> = conv.messages.iter().skip(floor).collect(); // 1) 反向扫描 assistant tool_calls,找最近一条同名同参调用 → 拿到旧 tool_call_id, // 同时累计同名同参调用总数(hit_count 供「第 N 次命中缓存」威慑)。 @@ -342,6 +351,29 @@ pub(crate) async fn find_cached_readonly_result( /// 只读缓存白名单已收敛为注册元数据(readonly 维度,DRY 债 #5): /// find_cached_readonly_result 改查 df_ai::ai_tools::is_readonly,本文件不再持名单。 +/// 写后失效:非只读工具成功回填后推进会话级缓存失效下界(readonly_cache_floor)。 +/// +/// 调用点 = 全部写工具真执行成功的回填处(Low 风险并行回填 / trust 放行回填 / +/// 边界放行回填 / ai_approve 审批后执行回填)。非只读工具(写文件/跑命令/改 DB 等) +/// 成功执行即视作世界状态变更,历史里的只读结果从此不可信——floor 推进到写结果 +/// push 之后,find_cached_readonly_result 扫描时跳过 floor 之前的全部旧调用(过界即脏)。 +/// +/// 单调不减(max):多写工具同轮/跨轮先后执行,floor 取最大(最后写),前段只读 +/// 结果全失效——保守方向安全(失效仅多一次真执行,误命中才喂脏数据)。 +/// +/// 须在持 session 锁的短临界区内调用(与写结果 push 同锁,防 floor 与 messages 漂移)。 +/// is_readonly 未注册的工具按非只读处理(保守推进 floor,防未知新写工具漏失效)。 +pub(crate) fn mark_readonly_cache_stale(conv: &mut super::super::PerConvState, tool_name: &str) { + crate::commands::ai::tool_registry::ensure_tool_meta_seeded(); + if df_ai::ai_tools::is_readonly(tool_name) { + return; // 只读工具不改世界状态,缓存照常有效 + } + let new_floor = conv.messages.len(); + if conv.readonly_cache_floor < new_floor { + conv.readonly_cache_floor = new_floor; + } +} + /// 递归对 JSON 对象的键做字典序排序（就地），数组成员也递归排序。 fn sort_object_keys(v: &mut serde_json::Value) { match v { @@ -545,7 +577,53 @@ mod tests { assert_ne!(a, b, "read_file 不同 offset 应不同 key(分段读正常)"); } - // ── AC-2 ③ 列举命令检测测试 ── + // ── 写后失效测试 ── + + /// mark_readonly_cache_stale:非只读工具推进 floor,只读工具不动。 + #[test] + fn mark_stale_advances_floor_for_write_tools_only() { + crate::commands::ai::tool_registry::ensure_tool_meta_seeded(); + let mut conv = super::super::super::PerConvState::new(); + // 三条历史消息(user + assistant + tool) + conv.messages.push(df_ai::provider::ChatMessage::user("hi")); + conv.messages.push(df_ai::provider::ChatMessage::assistant("ack")); + conv.messages.push(df_ai::provider::ChatMessage::tool_result("t1", "r")); + assert_eq!(conv.readonly_cache_floor, 0, "初值 0"); + + // read_file 只读:不推进 + mark_readonly_cache_stale(&mut conv, "read_file"); + assert_eq!(conv.readonly_cache_floor, 0, "只读工具不推进 floor"); + + // write_file 写:推进到当前 len(3) + mark_readonly_cache_stale(&mut conv, "write_file"); + assert_eq!(conv.readonly_cache_floor, 3, "写工具推进 floor 到 len"); + + // 再只读:不动(单调不减由 max 保证) + mark_readonly_cache_stale(&mut conv, "grep"); + assert_eq!(conv.readonly_cache_floor, 3); + + // run_command(写语义):取 max 不回退 + conv.messages.push(df_ai::provider::ChatMessage::user("again")); + mark_readonly_cache_stale(&mut conv, "run_command"); + assert_eq!(conv.readonly_cache_floor, 4, "第二个写工具推进到新 len"); + } + + /// floor 过滤逻辑纯函数级锁定:floor=min 钳制语义(消息缩短后 floor 不越界)。 + #[test] + fn readonly_cache_floor_clamped_by_len() { + let mut conv = super::super::super::PerConvState::new(); + conv.messages.push(df_ai::provider::ChatMessage::user("a")); + conv.messages.push(df_ai::provider::ChatMessage::user("b")); + mark_readonly_cache_stale(&mut conv, "write_file"); + assert_eq!(conv.readonly_cache_floor, 2); + // 模拟 pop_last_assistant_round/clear 缩短历史后:len=0,floor 语义钳到 0 + let floor = conv.readonly_cache_floor.min(conv.messages.len()); + conv.messages.clear(); + let floor2 = conv.readonly_cache_floor.min(conv.messages.len()); + assert_eq!(floor, 2, "len 未缩短时 floor 原样"); + assert_eq!(floor2, 0, "len 缩短后 floor 钳到 0(空扫描区间,安全)"); + } + #[test] fn listing_command_detects_powershell_and_unix() { diff --git a/src-tauri/src/commands/ai/augmentation/registry.rs b/src-tauri/src/commands/ai/augmentation/registry.rs index 6c7e00d..2135264 100644 --- a/src-tauri/src/commands/ai/augmentation/registry.rs +++ b/src-tauri/src/commands/ai/augmentation/registry.rs @@ -7,7 +7,7 @@ //! `resolve_all` 按 [`MentionRef::kind_str`] 分发到对应 Resolver,单条 resolve 失败 //! 收集进 errors 列表不阻断整批注入(返回 `(Vec, Vec)`)。 -use std::collections::HashMap; +use std::collections::{HashMap, HashSet}; use std::sync::Arc; use df_types::augmentation::{Augmentation, MentionRef, ResolveError}; @@ -45,6 +45,10 @@ impl ResolverRegistry { /// - 失败(Resolver 返 Err / kind 无对应 Resolver)→ push 进 `errors`,**不中断** /// 后续 mention(整批注入最大化保留可用上下文)。 /// + /// 同一引用去重(先到先得):`/技能` 前导与 @mention span 可指向同一技能,按 + /// `kind + ref_id`(skill 的 ref_id 即 name)跳过重复,避免同一正文重复 resolve + /// 并在 prompt 注入两份。首现位置决定保留哪条,顺序不变。 + /// /// 返回 `(成功列表, 失败列表)`,调用方据 errors 数决定是否提示用户某 mention 跳过。 pub async fn resolve_all( &self, @@ -53,7 +57,12 @@ impl ResolverRegistry { ) -> (Vec, Vec) { let mut augs = Vec::with_capacity(refs.len()); let mut errors = Vec::new(); + let mut seen: HashSet = HashSet::with_capacity(refs.len()); for rf in refs { + // 去重键 = kind + ref_id(id 或 skill 的 name):同键 mention 只 resolve 首条 + if !seen.insert(format!("{}:{}", rf.kind_str(), rf.ref_id())) { + continue; + } let kind = rf.kind_str(); match self.resolvers.get(kind) { Some(resolver) => match resolver.resolve(rf, loc).await { @@ -139,4 +148,48 @@ mod tests { assert!(augs.is_empty(), "两条 mention kind 均≠test 应全 KindMismatch"); assert_eq!(errors.len(), 2, "两条失败都应收集"); } + + /// 测试用 Resolver:kind 固定 "skill",resolve 固定返 Skill(模拟成功注入)。 + struct OkSkillResolver; + #[async_trait] + impl MentionResolver for OkSkillResolver { + fn kind(&self) -> &'static str { + "skill" + } + async fn resolve( + &self, + rf: &MentionRef, + _loc: ProviderLocality, + ) -> Result { + let name = match rf { + MentionRef::Skill { name } => name.clone(), + _ => String::new(), + }; + Ok(Augmentation::Skill { + name, + source: "test".into(), + body: "BODY".into(), + }) + } + } + + #[tokio::test] + async fn resolve_all_dedups_same_ref_first_wins() { + // 去重(先到先得):/技能 前导 + @mention span 指向同一技能时只 resolve/注入一份; + // 不同技能不受影响;首现位置决定保留顺序。 + let mut reg = ResolverRegistry::new(); + reg.register(Arc::new(OkSkillResolver)); + let refs = vec![ + MentionRef::Skill { name: "review".into() }, + MentionRef::Skill { name: "review".into() }, + MentionRef::Skill { name: "idea".into() }, + ]; + let (augs, errors) = reg.resolve_all(&refs, ProviderLocality::Local).await; + assert!(errors.is_empty()); + assert_eq!(augs.len(), 2, "同名技能两条应去重为一份"); + match &augs[0] { + Augmentation::Skill { name, .. } => assert_eq!(name, "review", "先到先得保留首条"), + other => panic!("expected Skill, got {:?}", other), + } + } } diff --git a/src/composables/ai/useAiSend.ts b/src/composables/ai/useAiSend.ts index 8df354d..a425a26 100644 --- a/src/composables/ai/useAiSend.ts +++ b/src/composables/ai/useAiSend.ts @@ -90,7 +90,9 @@ async function doSend(text: string, skill?: string, force = false, parts?: Conte state.messages.push({ id: userMsgId as MessageId, role: 'user', - content: text.trim(), + // 技能调用留痕(问题3,与后端 chat.rs user_content 同规则):带技能时本地气泡即刻补 + // /技能名 前缀,与落库文本一致——否则发送瞬间无痕、切回会话才见前缀,前后显示不一。 + content: skill ? `/${skill} ${text.trim()}` : text.trim(), // 仅在非空时挂 parts(纯文本消息保持 undefined,渲染走原 content 路径零回归) parts: parts && parts.length > 0 ? parts : undefined, // Input Augmentation: 仅在非空时挂 mentionSpans(供消息列表 chip 渲染;空则 undefined 零回归) diff --git a/src/i18n/en/aiTool.ts b/src/i18n/en/aiTool.ts index b8f599c..814c062 100644 --- a/src/i18n/en/aiTool.ts +++ b/src/i18n/en/aiTool.ts @@ -117,6 +117,8 @@ export default { // fetch_url header display name (useToolCardHeader): host-shortened prefix + fallback when no url fetchUrlPrefix: 'Fetch', fetchUrlFallback: 'Fetch URL', + // spawn_agent header display name (P3-3: task truncated to 25 chars, consumes backend "dispatch sub-agent" goal metadata) + spawnAgentPrefix: 'Spawn agent: {task}', installObscura: 'Install obscura via AI', obscuraInstallPrompt: 'Please install obscura (a Rust headless browser, used by fetch_url render=true for JS rendering). Try `cargo install obscura` first; if cargo is unavailable, download the prebuilt binary for this platform from the GitHub releases (github.com/h4ckf0r0day/obscura) and place it on PATH. After installing, verify with `obscura --version` and tell me I can retry the web fetch.', // grep cross-file content search results per output_mode phrasing diff --git a/src/i18n/zh-CN/aiTool.ts b/src/i18n/zh-CN/aiTool.ts index 1b87050..0c701d1 100644 --- a/src/i18n/zh-CN/aiTool.ts +++ b/src/i18n/zh-CN/aiTool.ts @@ -118,6 +118,8 @@ export default { // fetch_url 头部显示名(useToolCardHeader):host 精简前缀 + 无 url 时兜底 fetchUrlPrefix: '获取', fetchUrlFallback: '获取网页', + // spawn_agent 头部显示名(P3-3:任务书截 25 字,消费后端「派发子代理」目标元数据语义) + spawnAgentPrefix: '派发子代理: {task}', installObscura: 'AI 帮装 obscura', obscuraInstallPrompt: '请帮我安装 obscura(Rust 无头浏览器,用于 fetch_url 的 render=true JS 渲染)。优先尝试 `cargo install obscura`,若 cargo 不可用则从 GitHub releases(github.com/h4ckf0r0day/obscura)下载对应平台预编译二进制并放入 PATH。安装完成后用 `obscura --version` 验证,然后告诉我可以重试网页抓取了。', // grep 跨文件内容搜索结果按 output_mode 分文案