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

优化: 子代理体验与纪律注入

Browse Source 
- 工具纪律段抽共享同源注入
- 子代理轨迹浮层与完成卡身份
- 分离窗会话切换修复
- 960s 统一出口报告不再蒸发
This commit is contained in:
lxy
2026-10-04 16:27:31 +08:00
parent b2d44f85e9
commit 01d231c39d
12 changed files with 1150 additions and 152 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src-tauri/src/commands/ai/prompt.rs
+163 -63
View File
@@ -81,57 +81,68 @@ fn env_profile_date_volatile() -> String {
format!("今天是 {today}")
}
/// 按语言返回系统提示词的 (固定前缀, 项目上下文标题, 任务上下文标题)
fn system_prompt_parts(lang: &str) -> (&'static str, &'static str, &'static str) {
/// 按语言返回系统提示词的 (固定前缀, 项目上下文标题, 任务上下文标题)。
///
/// prefix 内的 read_symbol 优先与重复调用检测两条经 [`read_symbol_guidance`] /
/// [`duplicate_call_guidance`] 注入(与子代理工具纪律段同源,零复制)。
fn system_prompt_parts(lang: &str) -> (String, &'static str, &'static str) {
match lang {
"en" => (
"You are DevFlow's AI assistant. You can manage projects, tasks, ideas, and workflows, as well as analyze code, read/write files, and execute commands.\n\
Please respond in English.\n\n\
## Capabilities\n\
You can perform the following actions via tool calls:\n\
- Create/query projects, tasks, and ideas\n\
- Run workflows\n\
- Read file contents, list directories, create/write files\n\
- To understand/locate a code symbol (function/struct/class) prefer read_symbol(path, symbol): returns a skeleton (signature + line range + internal calls, minimal and high-density), avoiding read_file dumping the whole file into the prompt; pass full=true for the full definition body; read_symbol supports .rs/.ts/.tsx/.js/.jsx/.vue — fall back to read_file only for other languages or when the symbol is not found\n\
- Get file metadata (size, line count, binary detection)\n\
- Append content to files, search files by name pattern\n\n\
## Guidelines\n\
- Briefly explain your intent before executing actions\n\
- Ask for clarification if the user's intent is unclear\n\
- Prefer using tools to complete actions rather than just describing steps\n\
- When a tool call fails, clearly tell the user it failed and why. Never disguise a fallback action as the original intent's success (e.g. don't write to description to fake a directory binding), and never falsely report success\n\
- **Self-check before claiming completion**: Before announcing a task done, verify the result matches the original intent (grep key references / read back field values / cross-check). Fix discrepancies before declaring success; never claim all done without verification (repeatedly observed: missed reference renames, fields silently overwritten by later ops, forcing the user to manually verify)\n\
- **Duplicate call detection**: If a tool (e.g. read_file/search_files) has already been called with identical arguments and returned successfully, do not call it again — the result is already in the conversation history; review the context instead of re-executing\n\
## Focus\n\
- Always center your response on the core goal of the user's current request; the previous round's topic is only background, not the current task.\n\
- When the user switches topics (a clear new intent), follow the latest request; do not drag the old topic into the new answer.\n\
- Give the conclusion or action first, then add only necessary explanation; omit tangential information unrelated to the current request.\n\
- Do not proactively expand context (files/data) that is irrelevant to the current request.\n",
format!(
"You are DevFlow's AI assistant. You can manage projects, tasks, ideas, and workflows, as well as analyze code, read/write files, and execute commands.\n\
Please respond in English.\n\n\
## Capabilities\n\
You can perform the following actions via tool calls:\n\
- Create/query projects, tasks, and ideas\n\
- Run workflows\n\
- Read file contents, list directories, create/write files\n\
- {}\n\
- Get file metadata (size, line count, binary detection)\n\
- Append content to files, search files by name pattern\n\n\
## Guidelines\n\
- Briefly explain your intent before executing actions\n\
- Ask for clarification if the user's intent is unclear\n\
- Prefer using tools to complete actions rather than just describing steps\n\
- When a tool call fails, clearly tell the user it failed and why. Never disguise a fallback action as the original intent's success (e.g. don't write to description to fake a directory binding), and never falsely report success\n\
- **Self-check before claiming completion**: Before announcing a task done, verify the result matches the original intent (grep key references / read back field values / cross-check). Fix discrepancies before declaring success; never claim all done without verification (repeatedly observed: missed reference renames, fields silently overwritten by later ops, forcing the user to manually verify)\n\
- {}\n\
## Focus\n\
- Always center your response on the core goal of the user's current request; the previous round's topic is only background, not the current task.\n\
- When the user switches topics (a clear new intent), follow the latest request; do not drag the old topic into the new answer.\n\
- Give the conclusion or action first, then add only necessary explanation; omit tangential information unrelated to the current request.\n\
- Do not proactively expand context (files/data) that is irrelevant to the current request.\n",
read_symbol_guidance(lang),
duplicate_call_guidance(lang),
),
"\n## Current Projects\n",
"\n## Current Tasks\n",
),
_ => (
"你是 DevFlow 桌面应用的 AI 助手。你可以操作项目/任务/灵感/工作流，也可分析代码、读写文件、执行命令。\n\
必须使用简体中文回复，禁止使用繁体中文字符。\n\n\
## 当前能力\n\
你可以通过工具调用执行以下操作：\n\
- 创建/查询项目、任务、灵感\n\
- 运行工作流\n\
- 读取文件内容、列出目录、创建/写入文件\n\
- 理解/定位代码符号（函数/结构体/类等）优先用 read_symbol(path, symbol)：返回骨架（签名+行范围+内部调用，极小高密度），避免 read_file 全文回灌撑爆 prompt；需完整定义体传 full=true；read_symbol 支持 .rs/.ts/.tsx/.js/.jsx/.vue，其他语言或未找到符号时再退回 read_file\n\
- 获取文件元信息（大小、行数、二进制检测）\n\
- 追加写入文件、按名称模式搜索文件\n\n\
## 行为准则\n\
- 执行操作前简要说明你的意图\n\
- 如果不确定用户意图，先提问\n\
- 优先使用工具完成操作，而不是只描述步骤\n\
- 工具调用失败时必须明确告知用户失败原因，严禁用替代操作冒充原意图成功（如绑定目录失败不得改写描述冒充已绑定），也绝不谎报成功\n\
- **重复调用检测**：如果某个工具（如 read_file/search_files）已用相同参数调用过且返回成功，不要重复调用——会话历史里已有结果，回顾上下文而非重复执行\n\
## 聚焦准则\n\
- 始终围绕用户当前请求的核心目标回答；上一轮的主题只是背景，不是当前任务。\n\
- 用户切换话题（明显的新意图）时，以最新请求为准，不要把旧话题带进新回答。\n\
- 回答先给结论/动作，再补必要的解释；与当前请求无关的扩展信息省略。\n\
- 需要的上下文（文件/数据）若与当前请求无关，不要主动展开。\n",
format!(
"你是 DevFlow 桌面应用的 AI 助手。你可以操作项目/任务/灵感/工作流，也可分析代码、读写文件、执行命令。\n\
必须使用简体中文回复，禁止使用繁体中文字符。\n\n\
## 当前能力\n\
你可以通过工具调用执行以下操作：\n\
- 创建/查询项目、任务、灵感\n\
- 运行工作流\n\
- 读取文件内容、列出目录、创建/写入文件\n\
- {}\n\
- 获取文件元信息（大小、行数、二进制检测）\n\
- 追加写入文件、按名称模式搜索文件\n\n\
## 行为准则\n\
- 执行操作前简要说明你的意图\n\
- 如果不确定用户意图，先提问\n\
- 优先使用工具完成操作，而不是只描述步骤\n\
- 工具调用失败时必须明确告知用户失败原因，严禁用替代操作冒充原意图成功（如绑定目录失败不得改写描述冒充已绑定），也绝不谎报成功\n\
- {}\n\
## 聚焦准则\n\
- 始终围绕用户当前请求的核心目标回答；上一轮的主题只是背景，不是当前任务。\n\
- 用户切换话题（明显的新意图）时，以最新请求为准，不要把旧话题带进新回答。\n\
- 回答先给结论/动作，再补必要的解释；与当前请求无关的扩展信息省略。\n\
- 需要的上下文（文件/数据）若与当前请求无关，不要主动展开。\n",
read_symbol_guidance(lang),
duplicate_call_guidance(lang),
),
"\n## 当前项目\n",
"\n## 当前任务\n",
),
@@ -153,7 +164,56 @@ fn system_prompt_parts(lang: &str) -> (&'static str, &'static str, &'static str)
/// AI 看到 hint 即自主提示用户「装了能 JS 渲染 SPA/反爬」,征得同意后用 run_command 帮装。
///
/// 仅段加在 prefix 末尾(聚焦段后),不破坏现有结构。
fn doc_exploration_strategy_section(lang: &str) -> &'static str {
///
/// 段体拆为两块:核心纪律(URL 三工具分工/探索流程/禁臆造 URL/render=true 铁律)与
/// obscura 安装引导(涉用户交互与 run_command,父侧专用)。子代理工具纪律段
/// ([`sub_agent_tool_discipline_section`])只拼核心块。
fn doc_exploration_strategy_section(lang: &str) -> String {
format!(
"{}{}",
doc_strategy_core_section(lang),
doc_strategy_obscura_install_section(lang)
)
}
/// read_symbol 优先指引(中/英,不含行首 "- ")。父会话能力清单与子代理工具纪律段
/// 同源拼接(DRY,治指引真空:工具下发了纪律没下发,实测翻车坑在子代理原样复发)。
fn read_symbol_guidance(lang: &str) -> &'static str {
match lang {
"en" => "To understand/locate a code symbol (function/struct/class) prefer read_symbol(path, symbol): returns a skeleton (signature + line range + internal calls, minimal and high-density), avoiding read_file dumping the whole file into the prompt; pass full=true for the full definition body; read_symbol supports .rs/.ts/.tsx/.js/.jsx/.vue — fall back to read_file only for other languages or when the symbol is not found",
_ => "理解/定位代码符号（函数/结构体/类等）优先用 read_symbol(path, symbol)：返回骨架（签名+行范围+内部调用，极小高密度），避免 read_file 全文回灌撑爆 prompt；需完整定义体传 full=true；read_symbol 支持 .rs/.ts/.tsx/.js/.jsx/.vue，其他语言或未找到符号时再退回 read_file",
}
}
/// 重复调用检测指引(中/英,不含行首 "- ")。同源拼接见 read_symbol_guidance。
fn duplicate_call_guidance(lang: &str) -> &'static str {
match lang {
"en" => "**Duplicate call detection**: If a tool (e.g. read_file/search_files) has already been called with identical arguments and returned successfully, do not call it again — the result is already in the conversation history; review the context instead of re-executing",
_ => "**重复调用检测**：如果某个工具（如 read_file/search_files）已用相同参数调用过且返回成功，不要重复调用——会话历史里已有结果，回顾上下文而非重复执行",
}
}
/// 子代理工具纪律段(P1-1):与父会话 system prompt 同源拼接的工具使用纪律 ——
/// read_symbol 优先 / 重复调用检测 / URL 三工具纪律与禁臆造 URL / render=true 失败
/// 2 次必换方案 / 查配置用 get_app_config。子代理白名单缺省下发全量 Low 工具,
/// 纪律必须随工具同源下发,否则父会话实测蒸馏的防坑经验在子循环原样复发。
/// 裁剪不适用条款:obscura 安装引导(涉用户交互,子代理无交互面)、聚焦准则与
/// 用户交互类行为准则(执行前说明意图/向用户提问)。
pub(crate) fn sub_agent_tool_discipline_section(lang: &str) -> String {
format!(
"\n## 工具纪律\n\
- {}\n\
- {}\n{}{}",
read_symbol_guidance(lang),
duplicate_call_guidance(lang),
doc_strategy_core_section(lang),
app_config_query_guidance_section(lang),
)
}
/// 文档探索策略核心段(中/英):URL 三工具分工 + 嗅探→识别→提取→验证→应用流程 +
/// 禁臆造 URL + render=true 失败重试铁律。父会话与子代理同源拼接块。
fn doc_strategy_core_section(lang: &str) -> &'static str {
match lang {
"en" => "\n\
## Documentation Exploration Strategy\n\
@@ -189,17 +249,7 @@ fn doc_exploration_strategy_section(lang: &str) -> &'static str {
1. **Switch after 2 failures**: if `render=true` fails twice in a row (any failure signal) → **immediately stop retrying that URL**. Pick an alternative: (a) use `http_request` to hit the API endpoint directly; (b) ask the user to provide content/screenshots manually; (c) work with the data you already have and produce output now — don't wait for a perfect render.\n\
2. **Stop after the first success**: if the first `render=true` succeeds and returns markdown → **use the data you have immediately**; do NOT re-render the same URL to top up more content (what you got is what you use).\n\
3. **Never re-render the same URL**: once a URL's render result (success or failure) is known, never fire render=true at it again (history/cache is enough); for new content, switch URL or switch tactic.\n\
\n\
### JS-rendered docs & the obscura install hint\n\
`fetch_url` defaults to a static fetch (no JS). For SPA / anti-scrape / JS-rendered docs, the user (or you) can pass `render=true`. When obscura (the JS-rendering backend, github.com/h4ckf0r0day/obscura) is **not installed**, the result silently falls back to static and the returned JSON carries `render_mode: \"static_fallback\"` plus a `hint` field pointing to the install page. Watch the **`hint` field** specifically (not just `render_mode`) — it is the install signal:\n\
1. **Notice**: if the `fetch_url` result has `hint` set (or `render_mode` is `static_fallback`), obscura is missing and the page may be under-rendered.\n\
2. **Offer**: proactively tell the user — \"obscura is not installed; installing it lets fetch_url render JS-heavy SPA / anti-scrape docs. Want me to install it for you?\" Don't silently continue with the static fallback when the user clearly needs JS rendering.\n\
3. **Install on consent**: once the user agrees, run the install yourself via `run_command` and stream the output so the user sees progress — pick whichever path fits the user's environment:\n\
- Rust toolchain present: `cargo install obscura`\n\
- otherwise: download the matching prebuilt binary from github.com/h4ckf0r0day/obscura releases and place it on PATH\n\
(`run_command` is high-risk and requires user approval — the user's \"yes\" above is that approval; still announce the exact command before running.)\n\
4. **Re-fetch**: after install succeeds, tell the user \"obscura is ready\" and retry `fetch_url` with `render=true` to get the properly rendered markdown.\n\
This is a pure conversational guide — there is no UI button for it; you detect the hint and drive the install through dialogue.\n",
\n",
_ => "\n\
## 文档探索策略\n\
### 先选对 URL 工具(三者别混用)\n\
@@ -234,8 +284,28 @@ fn doc_exploration_strategy_section(lang: &str) -> &'static str {
1. **失败 2 次必换方案**:`render=true` 连续 2 次失败(任意失败信号) → **立即停止重试同一 URL**,改走以下任一:(a) 改用 `http_request` 直接打 API 端点;(b) 提示用户手动提供内容/截图;(c) 用已拿到的数据先整理输出,别等完美渲染。\n\
2. **第 1 次成功即收手**:`render=true` 第 1 次成功拿到 markdown → **立即用已有数据整理输出**,不要重复 render 同 URL 去「补充」更多内容(已拿即用)。\n\
3. **同 URL 不重复 render=true**:同一 URL 的 render 结果(成功或失败)已知,后续不再对它重复发 render=true(缓存/历史即用);需要新内容请换 URL 或换方案。\n\
\n\
### JS 渲染文档与 obscura 安装引导\n\
\n",
}
}
/// obscura 安装引导段(中/英,父侧专用):涉用户交互(征得同意帮装)与 run_command 高风险
/// 工具,子代理无用户交互面,不进子代理工具纪律段([`sub_agent_tool_discipline_section`])。
/// 与核心段拼接 = 原 [`doc_exploration_strategy_section`] 完整段(逐字节等价)。
fn doc_strategy_obscura_install_section(lang: &str) -> &'static str {
match lang {
// 注:核心段末尾已保留空行(…tactic.\n\n / …换方案。\n\n),本段自标题起拼,
// 拼接 = 原完整段逐字节等价,不再补前导换行。
"en" => "### JS-rendered docs & the obscura install hint\n\
`fetch_url` defaults to a static fetch (no JS). For SPA / anti-scrape / JS-rendered docs, the user (or you) can pass `render=true`. When obscura (the JS-rendering backend, github.com/h4ckf0r0day/obscura) is **not installed**, the result silently falls back to static and the returned JSON carries `render_mode: \"static_fallback\"` plus a `hint` field pointing to the install page. Watch the **`hint` field** specifically (not just `render_mode`) — it is the install signal:\n\
1. **Notice**: if the `fetch_url` result has `hint` set (or `render_mode` is `static_fallback`), obscura is missing and the page may be under-rendered.\n\
2. **Offer**: proactively tell the user — \"obscura is not installed; installing it lets fetch_url render JS-heavy SPA / anti-scrape docs. Want me to install it for you?\" Don't silently continue with the static fallback when the user clearly needs JS rendering.\n\
3. **Install on consent**: once the user agrees, run the install yourself via `run_command` and stream the output so the user sees progress — pick whichever path fits the user's environment:\n\
- Rust toolchain present: `cargo install obscura`\n\
- otherwise: download the matching prebuilt binary from github.com/h4ckf0r0day/obscura releases and place it on PATH\n\
(`run_command` is high-risk and requires user approval — the user's \"yes\" above is that approval; still announce the exact command before running.)\n\
4. **Re-fetch**: after install succeeds, tell the user \"obscura is ready\" and retry `fetch_url` with `render=true` to get the properly rendered markdown.\n\
This is a pure conversational guide — there is no UI button for it; you detect the hint and drive the install through dialogue.\n",
_ => "### JS 渲染文档与 obscura 安装引导\n\
`fetch_url` 默认走静态抓取(不跑 JS)。遇到 SPA / 反爬 / JS 渲染的文档,用户(或你)可传 `render=true`。当 obscura(JS 渲染后端,github.com/h4ckf0r0day/obscura)**未安装**时,结果会静默回退静态,返回的 JSON 带 `render_mode: \"static_fallback\"` 和一个指向安装页的 `hint` 字段。请**重点看 `hint` 字段**(而非只看 `render_mode`)——它就是安装信号:\n\
1. **察觉**:若 `fetch_url` 结果带 `hint`(或 `render_mode` 为 `static_fallback`),说明 obscura 缺失,页面可能没渲染完整。\n\
2. **提议**:主动告诉用户——「obscura 未安装,装了之后 fetch_url 就能渲染 JS 重的 SPA / 反爬文档。要我帮你装吗?」当用户明显需要 JS 渲染时,不要默默用静态回退糊弄过去。\n\
@@ -412,9 +482,9 @@ pub(crate) async fn build_system_prompt_with_excluded(
}
let (prefix, _, _) = system_prompt_parts(lang);
let mut prompt = env_profile_line();
prompt.push_str(prefix);
prompt.push_str(&prefix);
// 文档探索策略段(URL→fetch_url 嗅探→识别→提取→验证→应用,治盲目 http_request 瞎试)
prompt.push_str(doc_exploration_strategy_section(lang));
prompt.push_str(&doc_exploration_strategy_section(lang));
// 查 DevFlow 自身配置引导(治 AI 查配置绕 run_command PowerShell 内联脚本引号嵌套失败)
prompt.push_str(app_config_query_guidance_section(lang));
@@ -556,8 +626,8 @@ pub(crate) async fn build_system_prompt_parts(
// 易变段(日期 + 项目/任务清单)挪到消息流末尾,不失效 system 缓存。
let (prefix, _, _) = system_prompt_parts(lang);
let mut stable = env_profile_stable_body();
stable.push_str(prefix);
stable.push_str(doc_exploration_strategy_section(lang));
stable.push_str(&prefix);
stable.push_str(&doc_exploration_strategy_section(lang));
stable.push_str(app_config_query_guidance_section(lang));
stable.push_str(&build_custom_section(state).await);
@@ -847,6 +917,36 @@ mod tests {
assert!(en.contains("Do **NOT**"));
}
// P1-1 子代理工具纪律段:与父会话同源(read_symbol/重复调用检测/URL 纪律/查配置),
// 且裁剪掉用户交互条款(obscura 安装引导,子代理无交互面)。
#[test]
fn sub_agent_tool_discipline_section_shares_parent_text() {
let zh = sub_agent_tool_discipline_section("zh-CN");
assert!(zh.contains("## 工具纪律"));
assert!(zh.contains(read_symbol_guidance("zh-CN")), "read_symbol 指引须与父会话同源");
assert!(zh.contains(duplicate_call_guidance("zh-CN")), "重复调用检测须与父会话同源");
assert!(zh.contains("## 文档探索策略"), "URL 纪律核心段须注入");
assert!(zh.contains("严禁臆造 URL"));
assert!(zh.contains("失败 2 次必换方案"));
assert!(zh.contains("get_app_config"));
assert!(
!zh.contains("要我帮你装吗"),
"obscura 安装引导为用户交互条款,子代理不得注入: {}",
zh
);
assert!(!zh.contains("## 聚焦准则"), "会话面准则不进子代理");
// 父会话 prefix 与共享函数同字同源(DRY 锁:改共享函数父侧同步生效)
let (prefix, _, _) = system_prompt_parts("zh-CN");
assert!(prefix.contains(read_symbol_guidance("zh-CN")));
assert!(prefix.contains(duplicate_call_guidance("zh-CN")));
// 父侧完整文档探索段 = 核心 + obscura 引导(拆分不丢段)
let parent_doc = doc_exploration_strategy_section("zh-CN");
assert!(parent_doc.contains("失败 2 次必换方案"));
assert!(parent_doc.contains("要我帮你装吗"));
}
// 项目/任务清单注明语(中/英)存在
#[test]
fn listed_notes_present_zh_en() {