diff --git a/FEATURES.md b/FEATURES.md index 00e371d..887ef92 100644 --- a/FEATURES.md +++ b/FEATURES.md @@ -729,6 +729,7 @@ AnyBot 是一個桌面聊天 App，介面和運作方式照 Grok Bot 重做。 - 預設是這台 Mac 的 Docker（OrbStack 或 Docker Desktop）。 - 可以填 Docker 主機，例如 `ssh://me@server`（Tailscale 的機器名稱也行），Bot 的電腦就會在那台主機上；看畫面時 AnyBot 會自己開 SSH 通道。 +- **容器隔離**（設定 › 電腦 › 在哪裡執行）：標準（runc：容器和 Docker 主機共用核心，Chromium 關掉自己的沙盒，容器是唯一那道牆）或強隔離（gVisor／runsc：電腦有自己的虛擬核心，權限全拿掉只留開機要的幾個，Chromium 的沙盒打開，兩道牆）。Docker 主機沒有 gVisor 時強隔離不能選；設定是強隔離、主機卻沒有 gVisor 時（例如換了 Docker 主機）先用標準開機並寫進紀錄，不會整台開不起來。那一列顯示電腦實際用的是哪種，和設定不同時提示按「更新」。Mac 上用 `deploy/mac/gvisor-vm.sh` 建一台自帶 Docker＋gVisor 的 Lima 虛擬機（登入時自動啟動），`gvisor-migrate.sh` 把原本的檔案和登入搬過去（AnyBot 要先完全退出）。實測（Mac 上的 Lima VM）：開機、畫面、VNC、無障礙編號、xdotool、截圖都正常；13 個 Chromium 程序都沒有 --no-sandbox，網頁程序 seccomp=2、有自己的命名空間；每一步 exec 55–69 ms、截圖 95 ms。 - 換位置後會在新的地方重新建立；舊檔案可以先「儲存快照」，換過去再「從快照復原」。 - **「網路流量改走這台 Mac」**（原版：網路流量可以改走你的本機電腦，「僅對新連線生效」，管理員可以關掉）：打開後，Bot 的電腦之後執行的指令會經過這台 Mac 上的代理伺服器出去（例如要連公司內網或 VPN 才看得到的服務）。代理只聽這台 Mac 自己（127.0.0.1），用原版的埠 8791，被占用就往後找。設定頁顯示經過的連線數。Bot 的電腦在別的主機上時不能用。管理員可以用 `defaults write com.anybot.AnyBot AllowBoxTrafficViaThisMac false` 鎖住，這時顯示「🔒 由你的管理員設定，無法變更」。 @@ -1255,6 +1256,7 @@ computer_use_screenshot、click、mouse_move、mouse_button（按住／放開） - **記憶**：跟 Bot 說「記住我喜歡靠窗的座位」，它會用 remember 記下來；「忘記…」用 forget。每個 Bot 最多 200 條、每條 500 字（估計值），每次聊天都帶著。Bot 設定裡的「記得的事」可以一條條刪。 - **技能**：教 Bot 一套做法後請它「存成技能」（save_skill），所有 Bot 共用；系統指示只列名稱和一句說明，Bot 要用時先用 use_skill 讀完整做法。設定 →「連接器」頁的「技能」卡可以刪除。 +- **自己進化**：一件事做了 5 步以上（操作電腦、網頁、指令，不算記憶和設定），回答完後在背景回頭看一次：這是在網站或程式上動手操作的事（填表、發文、回留言、上傳下載、改設定），就把這次真的成功的最短做法存成技能（用畫面上的名字，不寫編號和座標；繞路和失敗寫成「注意」；使用者說過的偏好寫進去）；已經有同一件事的技能就合併成更好的一版，照著做而且很順就不動。回覆下面會多一行「（學會了新做法「…」）」或「（把做法「…」更新了）」。系統指示要 Bot 開始做事前先看技能清單、有符合的先 use_skill。實測（httpbin 訂 pizza，真的 Grok）：第一次 11 步學到技能；第二次先讀技能、10 步 24 秒一次做完，回頭看後判斷不用改。 - **範本**：側邊欄「＋」按右鍵 →「從範本建立」：研究助理、寫作夥伴、電腦助手、客服助理、開發助手，各有角色說明、指示、頭像和顏色。Bot 也能用 create_bot 的 template 參數建立。 ### 保險庫（照原版重做，2026-11-04） diff --git a/box/Dockerfile b/box/Dockerfile index 853b732..70c91a0 100644 --- a/box/Dockerfile +++ b/box/Dockerfile @@ -21,12 +21,20 @@ COPY overlay /etc/anybot/overlay COPY Xresources /etc/anybot/Xresources COPY menu /etc/anybot/menu RUN chmod +x /usr/local/bin/anybot-box-start /usr/local/bin/anybot-a11y /usr/local/bin/anybot-browser \ - # Chromium has no sandbox inside a container (no user namespaces); the box itself is the - # sandbox. Debian's chromium wrapper reads these flags for every launch (menu, bots, xdg-open). - && printf '%s\n' 'export CHROMIUM_FLAGS="$CHROMIUM_FLAGS --no-sandbox --disable-dev-shm-usage --no-first-run --no-default-browser-check --disable-gpu --password-store=basic --test-type"' > /etc/chromium.d/anybot + # Chromium has no sandbox inside a plain container (no user namespaces); the box itself is + # the sandbox. Under gVisor (容器隔離=runsc) the box has its own virtualized kernel and + # Chromium's sandbox works: AnyBot then starts the box with ANYBOT_BOX_SANDBOX=1 and the + # sandbox is switched back on (one more wall between a hostile web page and the machine). + # Debian's chromium wrapper reads these flags for every launch (menu, bots, xdg-open). + && printf '%s\n' \ + 'if [ "${ANYBOT_BOX_SANDBOX:-0}" = 1 ]; then' \ + ' export CHROMIUM_FLAGS="$CHROMIUM_FLAGS --disable-dev-shm-usage --no-first-run --no-default-browser-check --disable-gpu --password-store=basic"' \ + 'else' \ + ' export CHROMIUM_FLAGS="$CHROMIUM_FLAGS --no-sandbox --test-type --disable-dev-shm-usage --no-first-run --no-default-browser-check --disable-gpu --password-store=basic"' \ + 'fi' > /etc/chromium.d/anybot # Accessibility on for every app, including those started over `docker exec`. ENV DISPLAY=:1 SCREEN_SIZE=1280x800x24 GTK_MODULES=atk-bridge GTK_A11Y=atspi GNOME_ACCESSIBILITY=1 NO_AT_BRIDGE=0 EXPOSE 6080 -LABEL anybot.box.version="6" +LABEL anybot.box.version="7" ENTRYPOINT ["/usr/bin/tini", "--"] CMD ["/usr/local/bin/anybot-box-start"] diff --git a/box/browser b/box/browser index d78dee9..c4624d2 100644 --- a/box/browser +++ b/box/browser @@ -23,7 +23,12 @@ if [[ ! -e "$PROFILE" ]]; then ln -s /home/bot/.config/anybot-browser "$PROFILE" fi fi -FLAGS=(--no-sandbox --test-type --user-data-dir="$PROFILE" --remote-debugging-address=127.0.0.1 --remote-debugging-port=9222 +# The box (a container) is the sandbox by default, so Chromium runs --no-sandbox. Under gVisor +# (容器隔離=runsc) the box has a virtualized kernel of its own and AnyBot starts it with +# ANYBOT_BOX_SANDBOX=1 — Chromium's own sandbox works there, so keep it as a second wall. +SANDBOX=(--no-sandbox --test-type) +[ "${ANYBOT_BOX_SANDBOX:-0}" = 1 ] && SANDBOX=() +FLAGS=("${SANDBOX[@]}" --user-data-dir="$PROFILE" --remote-debugging-address=127.0.0.1 --remote-debugging-port=9222 --remote-allow-origins=* --force-renderer-accessibility --disable-blink-features=AutomationControlled --disable-dev-shm-usage --disable-gpu --use-angle=swiftshader --enable-unsafe-swiftshader --password-store=basic --no-first-run --no-default-browser-check --disable-session-crashed-bubble --hide-crash-restore-bubble diff --git a/deploy/mac/gvisor-migrate.sh b/deploy/mac/gvisor-migrate.sh new file mode 100755 index 0000000..23b09f7 --- /dev/null +++ b/deploy/mac/gvisor-migrate.sh @@ -0,0 +1,56 @@ +#!/usr/bin/env bash +# 把 Bot 的電腦整個搬到 gVisor VM：box 的全部檔案與瀏覽器登入都在 home 資料卷 +# （anybot-box-home，Chromium 的 profile 也在裡面），整卷複製過去後， +# AnyBot 在 VM 上掛同名資料卷開機＝檔案、登入、桌面設定全部跟著走。 +# 所有 Bot 共用這一台、這一份登入與檔案：不分帳號，資料夾照常共用。 +# +# 先完全退出 AnyBot 應用程式，然後： +# ./gvisor-migrate.sh 搬過去（本機這份保留，當備份） +# ./gvisor-migrate.sh --check 只比對兩邊，不動資料 +set -euo pipefail +VM=anybot-gvisor +VOL=anybot-box-home +VMHOST="unix://$HOME/.lima/$VM/sock/docker.sock" + +command -v docker >/dev/null 2>&1 || { echo "找不到 docker 指令"; exit 1; } +if [ -n "${DOCKER_HOST:-}" ]; then + echo "現在環境裡有 DOCKER_HOST=${DOCKER_HOST}（要從本機 Docker 搬出，請用普通終端機跑）"; exit 1 +fi +docker volume inspect "$VOL" >/dev/null 2>&1 || { echo "本機找不到資料卷 ${VOL}（box 還沒建立過？）"; exit 1; } +limactl list "$VM" --format '{{.Status}}' 2>/dev/null | grep -q Running || { echo "VM $VM 沒在跑，先跑 gvisor-vm.sh"; exit 1; } +# The Mac app's process is "AnyBot" and `anybot --serve` is "anybot": match either case. +if pgrep -ix anybot >/dev/null 2>&1; then + echo "AnyBot 還在跑：請先從選單完全退出（它會把 box 叫醒、資料卷會一直在用）。"; exit 1 +fi + +echo "本機 box 的檔案：$(docker run --rm -v "$VOL":/h alpine:3.20 sh -c 'find /h | wc -l') 個項目、$(docker run --rm -v "$VOL":/h alpine:3.20 sh -c 'du -sh /h | cut -f1')" + +if [ "${1:-}" = "--check" ]; then + echo "VM 上的檔案：$(DOCKER_HOST=$VMHOST docker run --rm -v "$VOL":/h alpine:3.20 sh -c 'find /h 2>/dev/null | wc -l') 個項目" + exit 0 +fi + +echo "停止本機的 box（容器留在本機當備份，不刪）…" +docker stop anybot-box anybot-box-home-holder >/dev/null 2>&1 || true + +echo "VM 上準備同名資料卷…" +DOCKER_HOST=$VMHOST docker volume create "$VOL" >/dev/null + +echo "搬移中（本機 → VM，同一條管線直送，不落地）…" +# 兩邊都是 Linux，檔案擁有者（uid 1000 = box 裡的 bot 使用者）直接對得上。 +docker run --rm -v "$VOL":/h -w /h alpine:3.20 tar -cf - . \ + | DOCKER_HOST=$VMHOST docker run --rm -i -v "$VOL":/h -w /h alpine:3.20 tar -xf - + +echo "核對…" +DOCKER_HOST=$VMHOST docker run --rm -v "$VOL":/h alpine:3.20 sh -c 'echo "VM 上的檔案：$(find /h | wc -l) 個項目、$(du -sh /h | cut -f1)"' + +cat </dev/null || true + echo "已移除 VM ${VM}。" + exit 0 +fi + +command -v limactl >/dev/null 2>&1 || { echo "先裝 Lima：brew install lima"; exit 1; } + +STATUS=$(limactl list "$VM" --format '{{.Status}}' 2>/dev/null || true) +if [ "$STATUS" = "Running" ]; then + echo "VM $VM 已在執行。" +elif [ -n "$STATUS" ]; then + echo "VM $VM 是 ${STATUS}，啟動中…" + limactl start --tty=false "$VM" +else + echo "建立 VM ${VM}（$CPUS CPU / ${MEM}GB / ${DISK}GB，第一次要下載映像檔，幾分鐘）…" + # template:docker-rootful = Lima's rootful Docker template (runsc wants a plain root daemon; + # the rootless docker template runs the daemon in a user namespace). + limactl start --name="$VM" --tty=false --cpus "$CPUS" --memory "$MEM" --disk "$DISK" template:docker-rootful +fi + +echo "VM 裡安裝 gVisor（runsc，systrap 平台：不需要 KVM，ARM VM 可用）…" +limactl shell "$VM" -- bash -ec ' +set -e +if ! command -v runsc >/dev/null 2>&1; then + sudo mkdir -p /etc/apt/keyrings + curl -fsSL https://gvisor.dev/archive.key | sudo gpg --dearmor -o /usr/share/keyrings/gvisor-archive-keyring.gpg + echo "deb [signed-by=/usr/share/keyrings/gvisor-archive-keyring.gpg] https://storage.googleapis.com/gvisor/releases release main" | sudo tee /etc/apt/sources.list.d/gvisor.list >/dev/null + sudo apt-get update -qq + sudo apt-get install -y -qq runsc +fi +# Register the runtime in the VM'\''s Docker daemon (merge, do not clobber other settings). +# Restart Docker only when that changed: a restart stops Bot 的電腦 running in this VM. +CHANGED=$(sudo python3 - </dev/null 2>&1 || limactl start-at-login "$VM" --enabled >/dev/null 2>&1 || echo "（沒能設定開機自動啟動：Mac 重開機後要再跑一次這個腳本）" + +# Lima forwards the VM's Docker socket to the Mac: AnyBot's Docker 主機用這個 unix:// 位址， +# 不用 ssh、不用碰 ~/.ssh/config。 +SOCK="$HOME/.lima/$VM/sock/docker.sock" +DOCKER_HOST="unix://$SOCK" docker info --format '{{range $k, $v := .Runtimes}}{{$k}} {{end}}' | grep -qw runsc || { echo "從 Mac 連到 VM 的 Docker 看不到 runsc"; exit 1; } +KERNEL=$(DOCKER_HOST="unix://$SOCK" docker run --rm --runtime=runsc --entrypoint cat alpine:3.20 /proc/version | cut -d' ' -f3) + +cat <, /// 示範教學 this bot has learned (for view_demo_step). pub demos: Vec, + /// The skills there are (a longer job is looked back on against them; see learn). + pub skills: Vec, } #[derive(Clone, Default)] @@ -103,6 +105,37 @@ fn review_action(job: &Job, review: &Review, action: &str, cancel: &AtomicBool) } } +/// 自己進化: after a longer job, the model looks back on its steps and keeps the way that worked +/// as a skill (or improves the one it followed), so the next time takes the short path. It runs +/// after Done: the person isn't kept waiting. +fn learn(job: &Job, history: &[Message], cancel: &AtomicBool, send: &dyn Fn(StreamEvent)) { + let Some(start) = history.iter().rposition(|m| m.role == Role::User && !m.steer) else { return }; + let turn = &history[start..]; + if !crate::memory::worth_learning(turn) { + return; + } + let req = ChatRequest { + system: crate::memory::LEARN_SYSTEM.into(), + history: vec![Message::user(&crate::memory::learn_brief(turn, &job.skills))], + tools: vec![], + ..job.req.clone() + }; + let out = match providers::run_step(&req, cancel, &|_| {}) { + Ok(o) => o, + Err(e) => { + eprintln!("AnyBot：整理做法時模型沒有回應：{e}"); + return; + } + }; + match (crate::memory::learned_skill(&out.text), turn.last()) { + (Some(k), Some(reply)) => { + eprintln!("AnyBot：學到做法「{}」", k.name); + send(StreamEvent::Learned(k, reply.at)); + } + _ => eprintln!("AnyBot：這次的做法沒有存：{}", out.text.split_whitespace().collect::>().join(" ").chars().take(200).collect::()), + } +} + /// Take what the person typed mid-turn and add it to the chat as steering. fn drain_steer(job: &Job, history: &mut Vec, send: &dyn Fn(StreamEvent)) -> bool { let queued: Vec = std::mem::take(&mut *job.steer.lock().unwrap()); @@ -121,6 +154,76 @@ fn target_for(job: &Job, args: &str) -> Result { job.computers.iter().find(|t| t.name == want).cloned().ok_or_else(|| not_here(job, &want)) } +/// One of the bot's own browsers (where, and which profile). +#[derive(Clone, PartialEq, Eq, Hash, Debug)] +enum BotBrowser { + Here(String), + Box(String), + Device(String, String), +} + +/// How many running jobs are using each bot browser (bots share the box's Chromium). +static BROWSER_JOBS: Mutex> = Mutex::new(Vec::new()); + +/// The bot's browsers a job used. When the job ends — answered, stopped or failed — each closes +/// the tabs the job left open and keeps the page it ended on (every browser_open is a new tab; +/// nothing used to close them, and the box ran out of memory with 19). A browser another job is +/// still using is left alone. +struct UsedBrowsers { + list: Vec, + via_name: String, +} + +/// A job starts using a bot browser. +fn claim(b: &BotBrowser) { + BROWSER_JOBS.lock().unwrap().push(b.clone()); +} + +/// A job is done with a bot browser: true when another job is still using it. +fn release(b: &BotBrowser) -> bool { + let mut jobs = BROWSER_JOBS.lock().unwrap(); + if let Some(i) = jobs.iter().position(|x| x == b) { + jobs.remove(i); + } + jobs.contains(b) +} + +impl UsedBrowsers { + fn add(&mut self, b: BotBrowser) { + if !self.list.contains(&b) { + claim(&b); + self.list.push(b); + } + } + + fn close_tabs(&mut self) { + for b in std::mem::take(&mut self.list) { + if release(&b) { + continue; + } + let args = |p: &str| if p.is_empty() { r#"{"action":"tidy"}"#.to_string() } else { serde_json::json!({"action": "tidy", "profile": p}).to_string() }; + let never = AtomicBool::new(false); + match &b { + BotBrowser::Here(p) => { + let _ = web::run_full("browser_tabs", &args(p), &never); + } + BotBrowser::Box(p) => { + let _ = crate::botbrowser::in_box(|| web::run_full("browser_tabs", &args(p), &never)); + } + BotBrowser::Device(host, p) => { + let _ = crate::remote::run_on_device(host, "browser_tabs", &args(p), &self.via_name); + } + } + } + } +} + +impl Drop for UsedBrowsers { + fn drop(&mut self) { + self.close_tabs(); + } +} + enum BrowserAt { Here, Box, @@ -542,6 +645,7 @@ pub fn spawn(job: Job, bot_id: u64, tx: Sender<(u64, StreamEvent)>, cancel: Arc< // The screen and browser steps of this turn (to stop the same step being tried over and over). let mut tried: Vec = Vec::new(); let mut local_on = job.local_execution; + let mut browsers = UsedBrowsers { list: vec![], via_name: job.via_name.clone() }; maybe_compact(&job, &mut history, COMPACT_THRESHOLD_PERCENT, &cancel, &send); for step in 0..MAX_STEPS { if step > 0 { @@ -593,6 +697,8 @@ pub fn spawn(job: Job, bot_id: u64, tx: Sender<(u64, StreamEvent)>, cancel: Arc< // Something typed while it answered keeps the same turn going. if job.steer.lock().unwrap().is_empty() { send(StreamEvent::Done); + browsers.close_tabs(); + learn(&job, &history, &cancel, &send); return; } continue; @@ -673,6 +779,14 @@ pub fn spawn(job: Job, bot_id: u64, tx: Sender<(u64, StreamEvent)>, cancel: Arc< let at = web::is_web_tool(&call.name).then(|| browser_place(&job, &call.name, &call.args)); if let Some((place, on)) = at.filter(|(place, _)| call.name == "web_fetch" || bots_browser(place, &call.name, &call.args)) { send(StreamEvent::ToolStatus(call.id.clone(), ToolStatus::Running)); + if call.name != "web_fetch" { + let profile = tools::args_of(&call.args).get("profile").and_then(|p| p.as_str()).unwrap_or("").trim().to_string(); + browsers.add(match &place { + BrowserAt::Here => BotBrowser::Here(profile), + BrowserAt::Box => BotBrowser::Box(profile), + BrowserAt::Device(host) => BotBrowser::Device(host.clone(), profile), + }); + } // 裝置連結: the AnyBot browser of another computer (its own logins). let (output, ok, img) = match place { BrowserAt::Device(host) => match crate::remote::run_on_device(&host, &call.name, &call.args, &job.via_name) { @@ -1284,6 +1398,7 @@ mod tests { steer: Arc::new(Mutex::new(vec![])), review: None, demos: vec![], + skills: vec![], } } @@ -1452,6 +1567,191 @@ mod tests { assert!(results[2].contains("盒子裡的測試頁 42"), "{}", results[2]); } + /// The box's browser works with the box's own disk (the brain runs elsewhere): an upload + /// finds the file there, a download and a PDF land there, and old tabs close. Needs Docker + /// and the box: `cargo test the_box_browser_uses_the_box_disk -- --ignored --nocapture`. + #[test] + #[ignore] + fn the_box_browser_uses_the_box_disk() { + let own = crate::model::BOT_COMPUTER_NAME; + let setup = r#"mkdir -p /tmp/anybot-files /tmp/anybot-up && cd /tmp/anybot-files && cat > index.html <<'EOF' +files

檔案測試

+ +

還沒選

下載報表 +EOF +printf 'a,b\n1,2\n' > report.csv && printf 'hello box' > '/tmp/anybot-up/上傳測試.txt' && (setsid python3 -m http.server 8768 --bind 127.0.0.1 >/dev/null 2>&1 &) ; sleep 1; echo served"#; + let check = "ls /home/bot/Downloads/report.csv /home/bot/Downloads/files-*.pdf && head -c 4 /home/bot/Downloads/files-*.pdf && echo"; + let pages = r#"curl -s 127.0.0.1:9222/json/list | python3 -c 'import json,sys; print("pages:", len([t for t in json.load(sys.stdin) if t["type"] == "page"]))'"#; + let clean = r#"pkill -f 'http[.]server 8768'; rm -rf /tmp/anybot-files /tmp/anybot-up /home/bot/Downloads/report.csv /home/bot/Downloads/files-*.pdf; curl -s 127.0.0.1:9222/json/list | python3 -c 'import json,sys,urllib.request; [urllib.request.urlopen("http://127.0.0.1:9222/json/close/" + t["id"]) for t in json.load(sys.stdin) if ":8768" in t["url"]]'; echo cleaned"#; + let mut replies: Vec<(&'static str, serde_json::Value)> = vec![ + ("shell", serde_json::json!({"command": setup})), + ("browser_open", serde_json::json!({"url": "http://127.0.0.1:8768/"})), + ("browser_upload", serde_json::json!({"path": "/tmp/anybot-up/上傳測試.txt"})), + ("browser_read", serde_json::json!({})), + ("browser_upload", serde_json::json!({"path": "/tmp/anybot-up/沒有這個.txt"})), + ("browser_download", serde_json::json!({"text": "下載報表"})), + ("browser_pdf", serde_json::json!({})), + ("shell", serde_json::json!({"command": check})), + ]; + for n in 1..=8 { + replies.push(("browser_open", serde_json::json!({"url": format!("http://127.0.0.1:8768/?n={n}")}))); + } + replies.push(("shell", serde_json::json!({"command": pages}))); + replies.push(("shell", serde_json::json!({"command": clean}))); + replies.push(("", serde_json::json!("好了"))); + let port = fake_model(replies); + let job = Job { + req: ChatRequest { kind: ProviderKind::Custom, base_url: format!("http://127.0.0.1:{port}/v1"), api_key: String::new(), model: "m".into(), system: String::new(), history: vec![Message::user("測試盒子裡的檔案")], tools: tools::specs(), session: String::new() }, + computers: vec![tools::Target { name: own.into(), place: tools::Place::Box { dir: crate::cloudbox::bot_dir(1) } }], + default_computer: own.into(), + roles: tools::Roles { own: own.into(), person: None, server: true }, + local_execution: false, + direct: Arc::new(AtomicBool::new(true)), + ..test_job() + }; + let (tx, rx) = channel(); + spawn(job, 1, tx, Arc::new(AtomicBool::new(false)), eframe::egui::Context::default()); + let mut results = vec![]; + loop { + match rx.recv_timeout(Duration::from_secs(180)).expect("loop stalled") { + (_, StreamEvent::ToolResult(m)) => { + println!("{}: {}\n", m.tool_name, m.text.chars().take(300).collect::()); + results.push(m.text) + } + (_, StreamEvent::Done) => break, + (_, StreamEvent::Error(e)) => panic!("{e}"), + _ => {} + } + } + assert!(results[0].contains("served"), "{}", results[0]); + assert!(results[1].contains("檔案測試"), "{}", results[1]); + assert!(results[2].contains("已把 /tmp/anybot-up/上傳測試.txt 放進上傳欄位，欄位裡現在是：上傳測試.txt（9 位元組）") && results[2].contains("已選：上傳測試.txt 9"), "{}", results[2]); + assert!(results[3].contains("已選：上傳測試.txt 9"), "{}", results[3]); + assert!(results[4].contains("這台電腦上沒有「/tmp/anybot-up/沒有這個.txt」"), "{}", results[4]); + assert!(results[5].contains("下載好了：/home/bot/Downloads/report.csv"), "{}", results[5]); + assert!(results[6].contains("已存成 PDF：/home/bot/Downloads/files-"), "{}", results[6]); + assert!(results[7].contains("%PDF") && !results[7].contains("No such file"), "{}", results[7]); + let count = results[16].split("pages:").nth(1).and_then(|n| n.trim().parse::().ok()).unwrap_or(99); + assert!(count <= 6, "{}", results[16]); + assert!(results[17].contains("cleaned"), "{}", results[17]); + } + + /// When a job ends, the tabs it left open in the box's browser close; the page it ended on + /// stays. Needs the box: `cargo test a_finished_job_closes_its_tabs -- --ignored --nocapture`. + #[test] + #[ignore] + fn a_finished_job_closes_its_tabs() { + let own = crate::model::BOT_COMPUTER_NAME; + let setup = "mkdir -p /tmp/anybot-tabs && cd /tmp/anybot-tabs && echo 'tab分頁' > index.html && (setsid python3 -m http.server 8770 --bind 127.0.0.1 >/dev/null 2>&1 &) ; sleep 1; echo served"; + let port = fake_model(vec![ + ("shell", serde_json::json!({"command": setup})), + ("browser_open", serde_json::json!({"url": "http://127.0.0.1:8770/?n=1"})), + ("browser_open", serde_json::json!({"url": "http://127.0.0.1:8770/?n=2"})), + ("browser_open", serde_json::json!({"url": "http://127.0.0.1:8770/?n=3"})), + ("", serde_json::json!("好了")), + ]); + let job = Job { + req: ChatRequest { kind: ProviderKind::Custom, base_url: format!("http://127.0.0.1:{port}/v1"), api_key: String::new(), model: "m".into(), system: String::new(), history: vec![Message::user("開三頁")], tools: tools::specs(), session: String::new() }, + computers: vec![tools::Target { name: own.into(), place: tools::Place::Box { dir: crate::cloudbox::bot_dir(1) } }], + default_computer: own.into(), + roles: tools::Roles { own: own.into(), person: None, server: true }, + local_execution: false, + direct: Arc::new(AtomicBool::new(true)), + ..test_job() + }; + let (tx, rx) = channel(); + spawn(job, 1, tx, Arc::new(AtomicBool::new(false)), eframe::egui::Context::default()); + loop { + match rx.recv_timeout(Duration::from_secs(120)).expect("loop stalled") { + (_, StreamEvent::Done) => break, + (_, StreamEvent::Error(e)) => panic!("{e}"), + _ => {} + } + } + std::thread::sleep(Duration::from_secs(2)); + let box_sh = |script: &str| String::from_utf8_lossy(&std::process::Command::new("docker").args(["exec", crate::cloudbox::container(), "sh", "-c", script]).output().unwrap().stdout).to_string(); + let left = box_sh(r#"curl -s 127.0.0.1:9222/json/list | python3 -c 'import json,sys; print("\n".join(t["url"] for t in json.load(sys.stdin) if t["type"] == "page"))'"#); + println!("left open:\n{left}"); + box_sh(r#"pkill -f 'http[.]server 8770'; rm -rf /tmp/anybot-tabs"#); + assert_eq!(left.trim(), "http://127.0.0.1:8770/?n=3", "{left}"); + } + + /// Words reach the control they name on a form laid out like httpbin's (each label and its + /// input inside a

): typing into 「Customer name:」, ticking 「Large」, pressing 「Submit + /// order」. They used to land on the

and do nothing. Needs the box: + /// `cargo test the_words_on_a_form_reach_its_controls -- --ignored --nocapture`. + #[test] + #[ignore] + fn the_words_on_a_form_reach_its_controls() { + let own = crate::model::BOT_COMPUTER_NAME; + let setup = r#"mkdir -p /tmp/anybot-form && cd /tmp/anybot-form && cat > index.html <<'EOF' +form +

+

+

+
Pizza Size +

+

+
Pizza Toppings +

+

+

+
+EOF +echo 'done送出了' > done.html; (setsid python3 -m http.server 8769 --bind 127.0.0.1 >/dev/null 2>&1 &) ; sleep 1; echo served"#; + let clean = r#"pkill -f 'http[.]server 8769'; rm -rf /tmp/anybot-form; curl -s 127.0.0.1:9222/json/list | python3 -c 'import json,sys,urllib.request; [urllib.request.urlopen("http://127.0.0.1:9222/json/close/" + t["id"]) for t in json.load(sys.stdin) if ":8769" in t["url"]]'; echo cleaned"#; + let port = fake_model(vec![ + ("shell", serde_json::json!({"command": setup})), + ("browser_open", serde_json::json!({"url": "http://127.0.0.1:8769/"})), + ("browser_type", serde_json::json!({"text": "Customer name:", "value": "李大華"})), + ("browser_type", serde_json::json!({"text": "Telephone", "value": "0922000111"})), + ("browser_click", serde_json::json!({"text": "Large"})), + ("browser_click", serde_json::json!({"text": "Extra Cheese"})), + ("browser_click", serde_json::json!({"text": "Onion"})), + ("browser_click", serde_json::json!({"text": "Submit order"})), + ("shell", serde_json::json!({"command": clean})), + ("", serde_json::json!("好了")), + ]); + let job = Job { + req: ChatRequest { kind: ProviderKind::Custom, base_url: format!("http://127.0.0.1:{port}/v1"), api_key: String::new(), model: "m".into(), system: String::new(), history: vec![Message::user("填表")], tools: tools::specs(), session: String::new() }, + computers: vec![tools::Target { name: own.into(), place: tools::Place::Box { dir: crate::cloudbox::bot_dir(1) } }], + default_computer: own.into(), + roles: tools::Roles { own: own.into(), person: None, server: true }, + local_execution: false, + direct: Arc::new(AtomicBool::new(true)), + ..test_job() + }; + let (tx, rx) = channel(); + spawn(job, 1, tx, Arc::new(AtomicBool::new(false)), eframe::egui::Context::default()); + let mut results = vec![]; + loop { + match rx.recv_timeout(Duration::from_secs(180)).expect("loop stalled") { + (_, StreamEvent::ToolResult(m)) => { + println!("{}: {}\n", m.tool_name, m.text.chars().take(240).collect::()); + results.push(m.text) + } + (_, StreamEvent::Done) => break, + (_, StreamEvent::Error(e)) => panic!("{e}"), + _ => {} + } + } + assert!(results[2].contains("欄位現在是「李大華」"), "{}", results[2]); + assert!(results[3].contains("欄位現在是「0922000111」"), "{}", results[3]); + let sent = &results[7]; + assert!(sent.contains("done.html?custname=%E6%9D%8E%E5%A4%A7%E8%8F%AF&custtel=0922000111&size=large&topping=cheese&topping=onion"), "{sent}"); + assert!(results[8].contains("cleaned"), "{}", results[8]); + } + + #[test] + fn a_browser_another_job_uses_keeps_its_tabs() { + let b = BotBrowser::Device("test-host-tabs".into(), String::new()); + claim(&b); + claim(&b); + assert!(release(&b), "the other job still uses it"); + assert!(!release(&b), "the last one out closes the tabs"); + assert!(!release(&b)); + } + #[test] fn the_same_miss_counts_as_the_same_step() { // Clicks a few pixels apart on the same strip are one step; another computer or target isn't. @@ -1596,6 +1896,7 @@ mod tests { steer: Arc::new(Mutex::new(vec![])), review: None, demos: vec![], + skills: vec![], }; let (tx, rx) = channel(); let cancel = Arc::new(AtomicBool::new(false)); @@ -1668,6 +1969,7 @@ mod tests { steer, review: None, demos: vec![], + skills: vec![], }; let (tx, rx) = channel(); spawn(job, 1, tx, Arc::new(AtomicBool::new(false)), eframe::egui::Context::default()); @@ -1746,6 +2048,7 @@ mod tests { steer: Arc::new(Mutex::new(vec![])), review: None, demos: vec![], + skills: vec![], }; let (tx, rx) = channel(); spawn(job, 1, tx, Arc::new(AtomicBool::new(false)), eframe::egui::Context::default()); @@ -1885,6 +2188,7 @@ mod tests { steer: Arc::new(Mutex::new(vec![])), review: None, demos: vec![], + skills: vec![], }; let (tx, rx) = channel(); spawn(job, 1, tx, Arc::new(AtomicBool::new(false)), eframe::egui::Context::default()); diff --git a/src/app.rs b/src/app.rs index 89515e6..96830c6 100644 --- a/src/app.rs +++ b/src/app.rs @@ -554,10 +554,20 @@ impl App { } } app.apply_theme(ctx); - // Real-data screenshots can open 設定 › 手機與帳號 (to press its buttons with ANYBOT_SHOT_CLICK). - if app.screenshot.is_some() && !app.demo && std::env::var("ANYBOT_SHOT_VIEW").as_deref() == Ok("remote") { - app.show_settings = true; - app.settings_tab = SettingsTab::Remote; + // Real-data screenshots can open 設定 › 手機與帳號 (to press its buttons with ANYBOT_SHOT_CLICK) + // or 設定 › 電腦 (容器隔離 shows what the box really runs on). + if app.screenshot.is_some() && !app.demo { + match std::env::var("ANYBOT_SHOT_VIEW").as_deref() { + Ok("remote") => { + app.show_settings = true; + app.settings_tab = SettingsTab::Remote; + } + Ok("computers") => { + app.show_settings = true; + app.settings_tab = SettingsTab::Computers; + } + _ => {} + } } // …or the side panel, or the computer viewer, as they look with that data. if app.screenshot.is_some() && !app.demo { @@ -832,6 +842,7 @@ impl App { steer, review, demos: b.demos.clone(), + skills: self.state.skills.clone(), }; agent::spawn(job, bot_id, self.tx.clone(), cancel, ctx.clone()); // (Its eyes start reading while it works; see face_pose.) @@ -1360,6 +1371,16 @@ impl App { fn drain_events(&mut self, ctx: &egui::Context) { while let Ok((id, ev)) = self.rx.try_recv() { + // 自己進化 arrives after the turn has ended: keep the skill, say so under the reply. + if let StreamEvent::Learned(k, reply_at) = ev { + let name = k.name.clone(); + let fresh = crate::memory::upsert_skill(&mut self.state.skills, k); + if let Some(m) = self.state.bot_mut(id).and_then(|b| b.messages.iter_mut().rev().find(|m| m.role == Role::Assistant && m.at == reply_at)) { + m.text.push_str(&crate::memory::learned_note(&name, fresh)); + } + self.dirty = true; + continue; + } // A stopped request may still deliver a few pieces; ignore them. if !self.streams.contains_key(&id) { continue; @@ -3560,6 +3581,20 @@ impl App { std::thread::sleep(std::time::Duration::from_millis(100)); } let secs = t0.elapsed().as_secs(); + // 自己進化 comes after the turn: wait for it when the turn was long enough to learn from. + let turn_msgs = app.state.bot(id).and_then(|b| b.messages.get(start..)).map(|m| m.to_vec()).unwrap_or_default(); + if crate::memory::worth_learning(&turn_msgs) { + let before = app.state.skills.clone(); + let t1 = Instant::now(); + while app.state.skills == before && t1.elapsed().as_secs() < 180 { + app.drain_events(ctx); + std::thread::sleep(std::time::Duration::from_millis(200)); + } + match app.state.skills.iter().find(|k| !before.contains(k)) { + Some(k) => println!("📚 學到（{} 秒）：{} — {}\n{}\n", t1.elapsed().as_secs(), k.name, k.description, k.instructions), + None => println!("📚 這次沒有存技能（{} 秒）", t1.elapsed().as_secs()), + } + } // The chat can shrink (compaction) between the snapshot and now: slice safely. let msgs = app.state.bot(id).and_then(|b| b.messages.get(start..)).map(|m| m.to_vec()).unwrap_or_default(); let clip = |s: &str, n: usize| -> String { @@ -3691,6 +3726,18 @@ impl eframe::App for App { } } } + // ANYBOT_SHOT_SCROLL="x,y,dy@frame": from that frame on, a wheel step of dy over x,y on + // each of 100 frames (egui eases a wheel in over real time; screenshot frames are quick). + if let Ok(sc) = std::env::var("ANYBOT_SHOT_SCROLL") { + if let Some((v, at)) = sc.split_once('@') { + let n: Vec = v.split(',').filter_map(|x| x.trim().parse().ok()).collect(); + let (at, f) = (at.trim().parse::().unwrap_or(30), self.frames + 1); + if n.len() == 3 && f >= at && f < at + 100 { + raw.events.push(egui::Event::PointerMoved(egui::pos2(n[0], n[1]))); + raw.events.push(egui::Event::MouseWheel { unit: egui::MouseWheelUnit::Point, delta: vec2(0.0, n[2]), phase: egui::TouchPhase::Move, source: egui::MouseWheelSource::Unknown, modifiers: Default::default() }); + } + } + } let Ok(spec) = std::env::var("ANYBOT_SHOT_CLICK") else { return }; for one in spec.split(';') { let (pos, right) = match one.strip_suffix(":right") { Some(p) => (p, true), None => (one, false) }; diff --git a/src/app/channels_ui.rs b/src/app/channels_ui.rs index 744fd39..81a3353 100644 --- a/src/app/channels_ui.rs +++ b/src/app/channels_ui.rs @@ -356,7 +356,7 @@ impl App { }); } setting_card(ui, p, |ui| { - setting_title(ui, "技能", "所有 Bot 共用的做法說明。在聊天裡教 Bot 一套做法並請它「存成技能」；Bot 需要時會先讀取再照做。", p); + setting_title(ui, "技能", "所有 Bot 共用的做法。Bot 做完較長的事，會自己把成功的做法整理成技能、下次照著做，做得更好時也會更新；你也可以在聊天裡教它一套做法。", p); if self.state.skills.is_empty() { hint(ui, "還沒有技能。", p); } diff --git a/src/app/computers_ui.rs b/src/app/computers_ui.rs index 5c1fed6..d1b77d3 100644 --- a/src/app/computers_ui.rs +++ b/src/app/computers_ui.rs @@ -36,6 +36,10 @@ pub struct ComputerUi { replace_secret: Option, secret_error: String, docker_host: Option, + /// 容器隔離 row: "runc" (standard) or "runsc" (gVisor). + box_runtime: String, + /// Checked in the background: the Docker host has gVisor, and the runtime the box runs on. + runtime_facts: Arc>>, show_log: bool, teach_name: String, teach_for: Option, @@ -51,6 +55,9 @@ pub struct ComputerUi { thumb_at: Option, } +/// (the Docker host has gVisor, the runtime the box runs on now) +type RuntimeFacts = (bool, Option); + fn status_color(r: &Run, p: Palette) -> Color32 { match r { Run::Running => Color32::from_rgb(0x14, 0xc3, 0x70), @@ -97,7 +104,11 @@ impl App { if !self.demo && self.cu.update_checked.map_or(true, |t| t.elapsed().as_secs() > 30) { self.cu.update_checked = Some(Instant::now()); let flag = self.cu.update_flag.clone(); - std::thread::spawn(move || flag.store(cloudbox::update_available(), Ordering::SeqCst)); + let facts = self.cu.runtime_facts.clone(); + std::thread::spawn(move || { + flag.store(cloudbox::update_available(), Ordering::SeqCst); + *facts.lock().unwrap() = Some((cloudbox::host_has_gvisor(), cloudbox::running_runtime())); + }); } self.cu.update_ready = self.cu.update_flag.load(Ordering::SeqCst); ui.spacing_mut().item_spacing.y = 0.0; @@ -616,6 +627,40 @@ impl App { } ui.add(egui::TextEdit::singleline(&mut edit).hint_text(t("這台電腦（預設）")).desired_width(180.0).font(egui::TextStyle::Monospace)); }); + let rt_now = cloudbox::box_runtime(); + let facts = self.cu.runtime_facts.lock().unwrap().clone(); + // What the box really runs on, and what pressing 更新 would change. + let status = match &facts { + Some((false, _)) if rt_now == "runsc" => "這台 Docker 主機沒有裝 gVisor：會用標準隔離開機".to_string(), + Some((_, Some(running))) if *running != rt_now => format!("目前是{}；按電腦的「更新」重建後改成{}", if running == "runsc" { "強隔離" } else { "標準" }, if rt_now == "runsc" { "強隔離" } else { "標準" }), + Some((_, Some(running))) => format!("目前是{}", if running == "runsc" { "強隔離（gVisor）" } else { "標準" }), + _ => String::new(), + }; + let gvisor_here = facts.as_ref().map(|f| f.0).unwrap_or(false); + sec_row_s(ui, p, "容器隔離", &status, "標準：容器和這台 Docker 主機共用 Linux 核心，Bot 開網頁時只有一層牆。Docker 主機裝了 gVisor 才能選強隔離：電腦得到自己的虛擬核心，Chromium 的沙盒也會打開，多一道牆。改完按電腦的「更新」重建後生效。", |ui| { + let rt = if self.cu.box_runtime.is_empty() { rt_now.clone() } else { self.cu.box_runtime.clone() }; + let std_on = rt == "runc"; + let gvis_on = rt == "runsc"; + // (The row lays out right to left: 強隔離 first puts 標準 on its left.) + let g = ui.add_enabled_ui(gvisor_here || gvis_on, |ui| ui.selectable_label(gvis_on, t("強隔離（gVisor）"))).inner; + if g.clicked() { + self.cu.box_runtime = "runsc".into(); + } + g.on_disabled_hover_text(t("這台 Docker 主機沒有裝 gVisor（runsc）")); + if ui.selectable_label(std_on, t("標準")).clicked() { + self.cu.box_runtime = "runc".into(); + } + if !self.cu.box_runtime.is_empty() && self.cu.box_runtime != rt_now { + if ui.add(primary_button("套用", p)).clicked() { + cloudbox::set_box_runtime(&self.cu.box_runtime); + self.cu.box_runtime.clear(); + self.cu.update_checked = None; + } + if ui.add(egui::Button::new(RichText::new(t("取消")).size(FONT_XS).color(p.text3)).frame(false)).clicked() { + self.cu.box_runtime.clear(); + } + } + }); }); if apply { cloudbox::set_docker_host(edit.trim()); diff --git a/src/appctl.rs b/src/appctl.rs index a1128d6..ba35341 100644 --- a/src/appctl.rs +++ b/src/appctl.rs @@ -128,6 +128,10 @@ pub fn describe(name: &str, args: &str) -> String { "delete_routine" | "run_routine_now" => format!("#{}", v.get("id").and_then(Value::as_u64).unwrap_or(0)), "message_bot" => format!("給 {}：{}", s("bot"), s("text").chars().take(40).collect::()), "create_bot" | "update_bot" => s("name").to_string() + &s("bot"), + // Which skill (the phone showed a bare 「刪除技能」), and which memory. + "save_skill" | "use_skill" | "delete_skill" => s("name"), + "remember" => s("fact").chars().take(40).collect(), + "forget" => s("what"), n if crate::pages::NAMES.contains(&n) => crate::pages::describe(n, &v), _ => String::new(), } @@ -425,11 +429,7 @@ pub fn apply(state: &mut State, bot_id: u64, name: &str, args: &str) -> Outcome if name.is_empty() || instructions.trim().is_empty() { return err("技能要有名稱和做法".into()); } - let k = crate::memory::Skill { name: name.clone(), description: s("description").unwrap_or_default(), instructions }; - match state.skills.iter_mut().find(|x| x.name == name) { - Some(x) => *x = k, - None => state.skills.push(k), - } + crate::memory::upsert_skill(&mut state.skills, crate::memory::Skill { name: name.clone(), description: s("description").unwrap_or_default(), instructions }); ok(format!("已存成技能「{name}」，所有 Bot 都能用。")) } "use_skill" => { @@ -598,6 +598,14 @@ pub fn apply(state: &mut State, bot_id: u64, name: &str, args: &str) -> Outcome mod tests { use super::*; + #[test] + fn a_step_on_skills_and_memories_names_which() { + assert_eq!(describe("delete_skill", r#"{"name":"本地活動報名"}"#), "本地活動報名"); + assert_eq!(describe("use_skill", r#"{"name":"Threads 發文"}"#), "Threads 發文"); + assert_eq!(describe("remember", r#"{"fact":"之後都用自己的電腦"}"#), "之後都用自己的電腦"); + assert_eq!(describe("forget", r#"{"what":"2"}"#), "2"); + } + #[test] fn chat_sets_up_and_manages_routines() { let mut st = State::samples(); diff --git a/src/botbrowser.rs b/src/botbrowser.rs index 68c9861..2bae05b 100644 --- a/src/botbrowser.rs +++ b/src/botbrowser.rs @@ -251,6 +251,14 @@ pub fn using_box() -> bool { } /// Start this profile's window if it isn't running. Returns (browser name, port, just started). +/// The debugging port of a bot browser that is already open (None when it isn't; never starts one). +pub fn running_port(profile: Option<&str>) -> Option { + if using_box() { + return crate::cloudbox::browser_running().then(crate::cloudbox::browser_port).and_then(Result::ok); + } + live_port(&resolve(profile).ok()?) +} + pub fn ensure(profile: &str, first_url: Option<&str>) -> Result<(&'static str, u16, bool), String> { if using_box() { let fresh = !crate::cloudbox::browser_running(); @@ -317,6 +325,7 @@ pub fn open(url: &str, profile: Option<&str>) -> Result { if r.status().as_u16() >= 400 { return Err(format!("{name} 沒辦法開新分頁（{}）", r.status())); } + close_old_tabs(port); } // Back when the page has loaded (not after a fixed 3 s: a slow page was still blank then), // with what it shows — the same as browser_read, one step fewer for every page. @@ -326,6 +335,23 @@ pub fn open(url: &str, profile: Option<&str>) -> Result { Ok(if page.is_empty() { head } else { format!("{head}\n\n{page}") }) } +/// Most tabs the bot's browser keeps. Every browser_open is a new tab and nothing closed them: +/// Bot 的電腦 once had 19 Threads pages open and its whole 2 GB in use (estimated). +const MAX_TABS: usize = 6; + +/// Close the tabs past MAX_TABS that were used longest ago (Chrome lists the latest first). +fn close_old_tabs(port: u16) { + let Ok(list) = local().get(&endpoint(port, "/json/list")).call().map_err(|e| e.to_string()).and_then(|r| r.into_body().read_json::().map_err(|e| e.to_string())) else { return }; + for id in old_tabs(&list) { + let _ = local().get(&endpoint(port, &format!("/json/close/{id}"))).call(); + } +} + +/// The page tabs past MAX_TABS in a /json/list answer (latest-used first). +fn old_tabs(list: &Value) -> Vec { + list.as_array().map(|a| a.iter().filter(|t| t["type"] == "page").skip(MAX_TABS).filter_map(|t| t["id"].as_str().map(String::from)).collect()).unwrap_or_default() +} + /// Longest wait for a page browser_open opened (estimated). const OPEN_WAIT_MS: u64 = 12_000; /// Text of the page handed back by browser_open (browser_read gives more when asked). @@ -431,6 +457,16 @@ mod tests { println!("{page}"); } + #[test] + fn only_the_tabs_used_longest_ago_close() { + let mut list: Vec = (0..9).map(|i| serde_json::json!({"type": "page", "id": format!("p{i}")})).collect(); + list.insert(2, serde_json::json!({"type": "browser_ui", "id": "ui"})); + list.insert(4, serde_json::json!({"type": "service_worker", "id": "sw"})); + assert_eq!(old_tabs(&Value::Array(list.clone())), vec!["p6", "p7", "p8"]); + list.truncate(5); + assert!(old_tabs(&Value::Array(list)).is_empty()); + } + #[test] fn urls_are_encoded_for_new_tabs() { assert_eq!(encode_component("https://www.threads.net/@me?x=1&y=2"), "https://www.threads.net/%40me%3Fx%3D1%26y%3D2"); diff --git a/src/browser.rs b/src/browser.rs index 027a0de..f6fcd98 100644 --- a/src/browser.rs +++ b/src/browser.rs @@ -228,7 +228,20 @@ fn find_js(r#ref: &str, text: &str) -> String { const vis = e => {{ const r = e.getBoundingClientRect(); return r.width > 1 && r.height > 1; }}; const all = [...document.querySelectorAll('a,button,[role],input,select,textarea,label,summary,li,td,th,span,div,p,img')].filter(vis); const hit = e => [t(e.innerText), t(e.getAttribute('aria-label')), t(e.getAttribute('placeholder')), t(e.title), t(e.value), e.labels && e.labels[0] ? t(e.labels[0].innerText) : '']; - return all.find(e => hit(e).includes(tx)) || all.filter(e => hit(e).some(h => h.includes(tx))).sort((a, b) => t(a.innerText).length - t(b.innerText).length)[0] || null; + // The control the words name, not the box around it: 「Customer name:」 is a

holding a + // label and its input, 「Submit order」 a

holding the button (its middle is empty page). + const ctl = 'a,button,input,select,textarea,summary,[role=button],[role=link],[role=checkbox],[role=radio],[role=switch],[role=tab],[role=menuitem],[role=option],[role=textbox],[role=combobox],[contenteditable=""],[contenteditable=true]'; + const isCtl = e => e.matches(ctl); + const inner = e => {{ + if (isCtl(e)) return e; + if (e.tagName === 'LABEL' && e.control && vis(e.control)) return e.control; + const c = [...e.querySelectorAll(ctl)].filter(vis); + return c.length === 1 ? c[0] : e; + }}; + const shortest = l => l.slice().sort((a, b) => t(a.innerText).length - t(b.innerText).length)[0]; + const pick = l => l.find(isCtl) || l.map(inner).find(isCtl) || shortest(l) || null; + const exact = all.filter(e => hit(e).includes(tx)); + return exact.length ? pick(exact) : pick(all.filter(e => hit(e).some(h => h.includes(tx)))); }})()"# ) } @@ -704,31 +717,53 @@ pub fn run(name: &str, args: &Value) -> Result<(String, Option), String> } "browser_upload" => { let path = s(args, "path"); - let full = crate::tools::expand_home(&path); - if !full.is_file() { + // The browser opens the file from its own disk: Bot 的電腦's when it runs there. + let in_box = botbrowser::using_box(); + let full = if in_box { crate::cloudbox::box_path(&path) } else { crate::tools::expand_home(&path).display().to_string() }; + let there = if in_box { crate::cloudbox::box_has_file(&full) } else { std::path::Path::new(&full).is_file() }; + if !there { return Err(format!("這台電腦上沒有「{path}」這個檔案")); } let js = format!("(() => {{ let el = {}; if (el && el.tagName !== 'INPUT') el = el.querySelector('input[type=file]') || el.closest('label')?.querySelector('input[type=file]'); return el || document.querySelector('input[type=file]'); }})()", find_js(&rf, &text)); let r = tab.call("Runtime.evaluate", json!({"expression": js, "returnByValue": false}))?; let obj = r["result"]["objectId"].as_str().ok_or("這一頁找不到上傳檔案的欄位")?.to_string(); - tab.call("DOM.setFileInputFiles", json!({"files": [full.display().to_string()], "objectId": obj}))?; + let before = snapshot(&mut tab); + tab.call("DOM.setFileInputFiles", json!({"files": [full], "objectId": obj}))?; tab.eval(&format!("(() => {{ const el = {js}; if (el) el.dispatchEvent(new Event('change', {{bubbles: true}})); return 1; }})()"))?; settle(1000); - Ok((format!("已把 {} 放進上傳欄位。", full.display()), None)) + // What the field holds now (a page that shows nothing about it left the bot + // checking with five more steps), and what the page did with it. + let held = tab.eval(&format!("(() => {{ const el = {js}; return el && el.files ? Array.from(el.files).map(f => f.name + '（' + f.size + ' 位元組）').join('、') : ''; }})()")).ok().and_then(|v| v.as_str().map(String::from)).unwrap_or_default(); + if held.is_empty() { + return Err(format!("放了 {full}，但上傳欄位還是空的（這一頁可能要先按「上傳」之類的按鈕，或欄位不是這一個）。")); + } + let after = snapshot(&mut tab); + let seen = changes(&before, &after); + let seen = if seen.contains("沒有變化") { String::new() } else { format!("\n{seen}") }; + Ok((format!("已把 {full} 放進上傳欄位，欄位裡現在是：{held}。{seen}"), None)) } "browser_download" => { - let dir = downloads_dir(); + // The file lands on the browser's own disk: Bot 的電腦's when it runs there. + let in_box = botbrowser::using_box(); + let dir = if in_box { crate::cloudbox::BOX_DOWNLOADS.to_string() } else { downloads_dir().display().to_string() }; + if in_box { + crate::cloudbox::box_mkdir(&dir)?; + } let clicked_at = std::time::SystemTime::now() - Duration::from_secs(1); - tab.call("Browser.setDownloadBehavior", json!({"behavior": "allow", "downloadPath": dir.display().to_string(), "eventsEnabled": false})).or_else(|_| tab.call("Page.setDownloadBehavior", json!({"behavior": "allow", "downloadPath": dir.display().to_string()})))?; + tab.call("Browser.setDownloadBehavior", json!({"behavior": "allow", "downloadPath": dir, "eventsEnabled": false})).or_else(|_| tab.call("Page.setDownloadBehavior", json!({"behavior": "allow", "downloadPath": dir})))?; let (x, y, what) = locate(&mut tab, &rf, &text)?; click_at(&mut tab, x, y, "left", 1)?; let start = Instant::now(); while start.elapsed().as_secs() < 60 { settle(700); // A new file, or one written again since the click (same name replaced). - let now: Vec = std::fs::read_dir(&dir).map(|r| r.flatten().filter(|e| e.metadata().and_then(|m| m.modified()).map(|t| t >= clicked_at).unwrap_or(false)).map(|e| e.path()).collect()).unwrap_or_default(); - if let Some(f) = now.iter().find(|p| !p.extension().map(|e| e == "crdownload" || e == "tmp").unwrap_or(false)) { - return Ok((format!("已點 {what}，下載好了：{}（可以用 read_file 讀它）", f.display()), None)); + let now: Vec = if in_box { + crate::cloudbox::box_new_files(&dir, clicked_at.duration_since(std::time::UNIX_EPOCH).map(|d| d.as_secs()).unwrap_or(0)) + } else { + std::fs::read_dir(&dir).map(|r| r.flatten().filter(|e| e.metadata().and_then(|m| m.modified()).map(|t| t >= clicked_at).unwrap_or(false)).map(|e| e.path().display().to_string()).collect()).unwrap_or_default() + }; + if let Some(f) = now.iter().find(|p| !(p.ends_with(".crdownload") || p.ends_with(".tmp"))) { + return Ok((format!("已點 {what}，下載好了：{f}（可以用 read_file 讀它）"), None)); } } Err(format!("已點 {what}，但 60 秒內沒有下載到檔案（有些網站要先選格式或日期，或在新分頁產生報表）。")) @@ -739,9 +774,18 @@ pub fn run(name: &str, args: &Value) -> Result<(String, Option), String> let bytes = base64::engine::general_purpose::STANDARD.decode(r["data"].as_str().unwrap_or("")).map_err(|e| e.to_string())?; let (title, _, _) = snapshot(&mut tab); let safe: String = title.chars().filter(|c| !"/\\:*?\"<>|".contains(*c)).take(60).collect(); - let path = downloads_dir().join(format!("{}-{}.pdf", if safe.trim().is_empty() { "網頁".into() } else { safe }, chrono::Local::now().format("%Y%m%d-%H%M%S"))); - std::fs::write(&path, bytes).map_err(|e| e.to_string())?; - Ok((format!("已存成 PDF：{}", path.display()), None)) + let name = format!("{}-{}.pdf", if safe.trim().is_empty() { "網頁".into() } else { safe }, chrono::Utc::now().with_timezone(&crate::routines::person_tz()).format("%Y%m%d-%H%M%S")); + // Saved where the bot can open it: on Bot 的電腦 when that's the browser's computer. + let path = if botbrowser::using_box() { + let p = format!("{}/{name}", crate::cloudbox::BOX_DOWNLOADS); + crate::cloudbox::box_write(&p, &bytes)?; + p + } else { + let p = downloads_dir().join(name); + std::fs::write(&p, bytes).map_err(|e| e.to_string())?; + p.display().to_string() + }; + Ok((format!("已存成 PDF：{path}"), None)) } "browser_dialog" => { let mut p = json!({"accept": args["accept"].as_bool().unwrap_or(true)}); @@ -789,7 +833,19 @@ fn shrink(b64: &str) -> Option { Some(base64::engine::general_purpose::STANDARD.encode(out.into_inner())) } +/// A job is done with this browser: keep the page it ended on, close the rest (the list is +/// latest-used first). A browser that isn't open stays closed. Not offered to the model; the +/// loop calls it (browser_tabs action "tidy") when a job ends. +fn tidy(profile: Option<&str>) -> Result { + let Some(port) = botbrowser::running_port(profile) else { return Ok("瀏覽器沒有開著。".into()) }; + let closed = targets(port)?.iter().skip(1).filter(|t| botbrowser::local().get(&format!("http://127.0.0.1:{port}/json/close/{}", t["id"].as_str().unwrap_or(""))).call().is_ok()).count(); + Ok(format!("關了 {closed} 個用完的分頁。")) +} + fn tabs(args: &Value, profile: Option<&str>) -> Result { + if s(args, "action") == "tidy" { + return tidy(profile); + } let (_, port) = port_of(profile)?; let http = |path: &str, put: bool| -> Result { let url = format!("http://127.0.0.1:{port}{path}"); diff --git a/src/cloudbox.rs b/src/cloudbox.rs index 99445f6..cb32e1e 100644 --- a/src/cloudbox.rs +++ b/src/cloudbox.rs @@ -167,6 +167,13 @@ pub struct Config { /// loopback-only, but the box's own programs can reach it, and they should need a key). #[serde(default)] pub vnc_password: String, + /// 容器隔離: the Docker runtime the box runs on. Empty = Docker's default (runc: the box + /// shares the host's kernel, and Chromium runs with its own sandbox off). On a Linux Docker + /// host with gVisor installed, "runsc" gives the box a virtualized kernel and Chromium's + /// sandbox comes back — a second wall for the pages bots open. Takes effect on the next + /// 更新/重設/復原 (the container is recreated). + #[serde(default)] + pub runtime: String, } /// A scheduled update that was missed may still run for this long (the original: one hour). @@ -312,6 +319,39 @@ fn save_config(c: &Config) { let _ = std::fs::write(config_path(), serde_json::to_string_pretty(c).unwrap_or_default()); } +/// 容器隔離: the box's Docker runtime — "runc" (the default: a container sharing the host's +/// kernel) or "runsc" (gVisor: a virtualized kernel and user-space seccomp, for Docker hosts +/// that have it installed). Anything else falls back to the default. +pub fn box_runtime() -> String { + match config().lock().unwrap().runtime.as_str() { + "runsc" => "runsc".into(), + _ => "runc".into(), + } +} + +/// The box's Docker host has gVisor (a runtime registered as runsc). +pub fn host_has_gvisor() -> bool { + dk(&["info", "--format", "{{range $k, $v := .Runtimes}}{{$k}} {{end}}"]).map(|o| o.split_whitespace().any(|r| r == "runsc")).unwrap_or(false) +} + +/// The runtime the box is running on now ("runc", "runsc"), when there is a box. +pub fn running_runtime() -> Option { + dk(&["inspect", "-f", "{{.HostConfig.Runtime}}", container()]).ok().map(|r| r.trim().to_string()).filter(|r| !r.is_empty()) +} + +/// Changing it takes effect on the box's next rebuild (更新/重設/復原), where the container +/// is created again — nothing is recreated here behind the person's back. +pub fn set_box_runtime(r: &str) { + let r = if r.trim() == "runsc" { "runsc" } else { "" }; + let mut c = config().lock().unwrap(); + if c.runtime == r { + return; + } + c.runtime = r.to_string(); + save_config(&c); + log(&format!("容器隔離改為「{}」，下次重建電腦（更新、重設或復原）後生效。", if r == "runsc" { "gVisor（虛擬核心）" } else { "標準（共用主機核心）" })); +} + pub fn egress_via_mac() -> bool { config().lock().unwrap().egress_via_mac } @@ -380,7 +420,8 @@ fn maybe_scheduled_update() { (c.auto_update_at.clone(), c.auto_update_done.clone()) }; let Some(at) = at else { return }; - let now = chrono::Local::now(); + // The person's clock (a brain on a server keeps UTC: 03:00 used to be 11:00 in Taipei). + let now = chrono::Utc::now().with_timezone(&crate::routines::person_tz()); let today = now.format("%Y-%m-%d").to_string(); if done == today || !in_schedule_window(&at, now.time()) || is_busy() || !update_available() { return; @@ -688,8 +729,40 @@ fn create_container() -> Result<(), String> { let env = format!("SCREEN_SIZE={SCREEN}"); let vnc = format!("VNC_PASSWORD={}", vnc_password()); log("正在開一台新的電腦…"); + // 容器隔離: on a Docker host with gVisor (`runsc`), the box gets a virtualized kernel + // instead of the host's, its seccomp filters run in user space, and Chromium's own sandbox + // works inside the box (ANYBOT_BOX_SANDBOX=1 re-enables it; see box/Dockerfile). With the + // plain runc runtime the box itself is the sandbox, like the original. + // A Docker host without gVisor (the setting moved with a new Docker host, or gVisor was + // removed) would leave no computer at all ("unknown runtime"): open it the standard way and + // say so; the settings page shows which one it runs on. + let hardened = box_runtime() == "runsc" && host_has_gvisor(); + if box_runtime() == "runsc" && !hardened { + log("這台 Docker 主機沒有 gVisor（runsc），電腦先用標準隔離開機；裝好 gVisor 後按「更新」重建。"); + } + let runtime = "runsc"; + let sandbox_env = "ANYBOT_BOX_SANDBOX=1"; + let caps: Vec = ["CHOWN", "SETUID", "SETGID", "FOWNER", "DAC_OVERRIDE", "KILL", "NET_BIND_SERVICE"] + .iter() + .flat_map(|c| ["--cap-add".to_string(), (*c).to_string()]) + .collect(); + let mut args: Vec<&str> = vec!["run", "-d", "--name", container(), "--hostname", "bot-computer", "--shm-size", "1g", "--pids-limit", "4096", "--memory", "2g", "--cpus", "2", "--add-host", "host.docker.internal:host-gateway", "-p", &publish, "-v", &vol, "-e", &env, "-e", &vnc]; + if hardened { + // (No swap on top of the 2 GB under gVisor; a standard box keeps Docker's default.) + args.extend(["--memory-swap", "2g"]); + args.push("--runtime"); + args.push(runtime); + args.push("--cap-drop"); + args.push("ALL"); + for c in &caps { + args.push(c); + } + args.push("-e"); + args.push(sandbox_env); + } + args.extend(["--label", "anybot.box=1", IMAGE]); // host.docker.internal → this Mac (for 網路流量改走你的電腦; Docker Desktop and OrbStack already have it). - dk(&["run", "-d", "--name", container(), "--hostname", "bot-computer", "--shm-size", "1g", "--pids-limit", "4096", "--memory", "2g", "--cpus", "2", "--add-host", "host.docker.internal:host-gateway", "-p", &publish, "-v", &vol, "-e", &env, "-e", &vnc, "--label", "anybot.box=1", IMAGE]).map_err(|e| format!("開不了電腦：{}", tail(&e, 400)))?; + dk(&args).map_err(|e| format!("開不了電腦：{}", tail(&e, 400)))?; ensure_holder(); Ok(()) } @@ -1598,6 +1671,46 @@ fn box_exec(script: &str) -> Result { } } +// ---------- files for the box's browser ---------- +// The brain drives the box's Chromium from outside, but that browser reads and writes the box's +// own disk: an upload, a download or a PDF has to be looked for there, not on the brain (a +// cloud brain once told the bot an attachment sitting in the box "isn't on this computer", and +// the bot spent 24 steps getting one picture into a post). + +/// Where the box's browser keeps downloads and PDFs (the bot reads them there). +pub const BOX_DOWNLOADS: &str = "/home/bot/Downloads"; + +/// A path on Bot 的電腦 as written by the bot: `~` is the bot user's home. +pub fn box_path(p: &str) -> String { + match p.trim() { + "~" => "/home/bot".into(), + t => t.strip_prefix("~/").map(|rest| format!("/home/bot/{rest}")).unwrap_or_else(|| t.to_string()), + } +} + +/// There is a file at `path` on Bot 的電腦. +pub fn box_has_file(path: &str) -> bool { + box_exec(&format!("[ -f {} ] && echo yes", crate::tools::sh_quote(path))).map(|o| o.trim() == "yes").unwrap_or(false) +} + +/// Files in `dir` on Bot 的電腦 written at or after `since` (seconds since 1970). +pub fn box_new_files(dir: &str, since: u64) -> Vec { + let script = format!("find {} -maxdepth 1 -type f -newermt @{since} 2>/dev/null; true", crate::tools::sh_quote(dir)); + box_exec(&script).map(|o| o.lines().map(str::trim).filter(|l| !l.is_empty()).map(String::from).collect()).unwrap_or_default() +} + +/// Make a folder on Bot 的電腦 (as the bot user). +pub fn box_mkdir(dir: &str) -> Result<(), String> { + box_exec(&format!("mkdir -p {}", crate::tools::sh_quote(dir))).map(|_| ()) +} + +/// Write a file on Bot 的電腦 (as the bot user), making its folder first. +pub fn box_write(path: &str, bytes: &[u8]) -> Result<(), String> { + let dir = std::path::Path::new(path).parent().map(|d| d.display().to_string()).unwrap_or_else(|| "/home/bot".into()); + let script = format!("mkdir -p {} && cat > {}", crate::tools::sh_quote(&dir), crate::tools::sh_quote(path)); + dk_with_input(&["exec", "-i", "-u", "bot", container(), "sh", "-c", &script], Some(bytes), 60_000).map(|_| ()).map_err(|e| format!("存不進 Bot 的電腦：{e}")) +} + /// The numbered controls of the box's latest look (centres in desktop pixels). static BOX_SPOTS: Mutex> = Mutex::new(None); /// The latest look's fingerprint (to say when a step changed nothing). @@ -2457,6 +2570,13 @@ mod tests { let _ = run_on("shell", r#"{"command":"cd .. && rm -rf bot-998"}"#, &t, &cancel, &ids); } + #[test] + fn box_paths_are_the_bot_users() { + assert_eq!(box_path("~/bot-6/附件/IMG_0587.jpeg"), "/home/bot/bot-6/附件/IMG_0587.jpeg"); + assert_eq!(box_path(" ~ "), "/home/bot"); + assert_eq!(box_path("/tmp/a.jpg"), "/tmp/a.jpg"); + } + #[test] fn schedule_window_is_one_hour() { let t = |s: &str| chrono::NaiveTime::parse_from_str(s, "%H:%M").unwrap(); @@ -2466,7 +2586,11 @@ mod tests { assert!(!in_schedule_window("03:00", t("02:59"))); assert!(in_schedule_window("23:30", t("00:10"))); assert!(!in_schedule_window("bad", t("00:10"))); - assert_eq!(wanted_version(), "6"); + // On a server (UTC), 03:00 is the person's 03:00: 19:00 UTC in Taipei. + use chrono::TimeZone; + let at = chrono::Utc.with_ymd_and_hms(2026, 11, 6, 19, 0, 0).unwrap().with_timezone(&chrono_tz::Asia::Taipei); + assert!(in_schedule_window("03:00", at.time())); + assert_eq!(wanted_version(), "7"); } #[test] diff --git a/src/memory.rs b/src/memory.rs index 7b9db8b..30707f3 100644 --- a/src/memory.rs +++ b/src/memory.rs @@ -63,6 +63,121 @@ pub fn forget(list: &mut Vec, what: &str) -> Result { } } +/// Add a skill, or replace the one with the same name. True when it is new. +pub fn upsert_skill(list: &mut Vec, k: Skill) -> bool { + match list.iter_mut().find(|x| x.name == k.name) { + Some(x) => { + *x = k; + false + } + None => { + list.push(k); + true + } + } +} + +// ---------- 自己進化: a job's way that worked becomes a skill ---------- +// The bot forgot everything between chats: a post with a picture on Threads took it 39 steps, +// and the next one would have taken as many. After a longer job the model looks back at the +// steps and writes down the short path that worked (or improves the skill it followed); the +// next time it reads that first. + +/// Steps on computers and pages a job needs before it is worth looking back on (estimated). +pub const LEARN_MIN_STEPS: usize = 5; +/// Longest learned skill (estimated; a skill is read whole before each such job). +const MAX_SKILL_CHARS: usize = 4000; + +/// A turn that did enough real work, and didn't already save a skill itself. +pub fn worth_learning(turn: &[crate::model::Message]) -> bool { + let calls: Vec<&str> = turn.iter().flat_map(|m| m.tool_calls.iter().map(|c| c.name.as_str())).collect(); + !calls.contains(&"save_skill") && calls.iter().filter(|n| !crate::appctl::is_app_tool(n) && **n != "react").count() >= LEARN_MIN_STEPS +} + +pub const LEARN_SYSTEM: &str = "你在幫一個會操作電腦的 Bot 整理經驗，讓它下次做同樣的事更快、不再犯同樣的錯。下面是它剛做完的一件事：使用者的要求、每一步和結果、最後的回覆，還有它已經有的技能。 + +先判斷要不要存。使用者希望它從做過的事學會、下次不用重新摸索：只要這件事是在某個網站或程式上動手操作（填表、發文、回留言、按讚、上傳下載、改設定、跑一套固定的檢查），就要存，就算這次只做了一次。只有閒聊、單純回答問題、只是查資料讀內容，才不用存。這次沒做成功的，不要存成做法；但如果已經有這件事的技能，可以把「這樣做行不通」補進去。 + +要存的話，寫一份「下次照著做就會成功」的技能： +- 寫這次真的成功的那條最短路徑，一步一行，寫出用哪個工具、在哪台電腦、哪個網址或程式（每一步結果開頭 [ ] 裡的名字，就是那一步用的電腦）。 +- 繞路和失敗的嘗試不要寫成步驟；把「哪裡會卡、要怎麼避開」寫成「注意：」。 +- 用畫面上看得到的名字指東西（例如 browser_click 的 text 寫「Attach media」、browser_type 的 text 寫欄位名「Customer name」），不要寫元素編號或座標，那些每次都會變。browser_open 的結果就有這一頁的元素清單，不用另外截圖。 +- 寫通用的做法，不要寫這次的內容（這篇文的文字、這張圖的檔名）；使用者表達過的偏好（語氣、格式、一定要附圖、用哪台電腦）要寫進去。 +- 已經有做同一件事的技能，就用同一個名稱，把舊的和這次學到的合併成更好的一版：舊的對的保留，錯的改掉。這次照著技能做而且很順，就不用再存。 + +只回一個 JSON，不要別的字： +不用存：{\"save\": false, \"why\": \"一句話說為什麼\"} +要存：{\"save\": true, \"name\": \"短名稱\", \"description\": \"一行：什麼時候用\", \"instructions\": \"做法\"}"; + +fn cut(s: &str, n: usize) -> String { + let flat = s.split_whitespace().collect::>().join(" "); + if flat.chars().count() <= n { flat } else { format!("{}…", flat.chars().take(n).collect::()) } +} + +/// What the model looks back on: the request, each step with its result, the answer and the +/// skills there are. +pub fn learn_brief(turn: &[crate::model::Message], skills: &[Skill]) -> String { + use crate::model::Role; + let mut steps: Vec = vec![]; + for (i, m) in turn.iter().enumerate() { + if m.role == Role::User && i > 0 { + steps.push(format!("（使用者中途說：{}）", cut(&m.text, 300))); + } + for c in &m.tool_calls { + let result = turn.iter().find(|r| r.role == Role::Tool && r.tool_call_id == c.id).map(|r| r.text.as_str()).unwrap_or(""); + let failed = if c.status == crate::model::ToolStatus::Failed { "（失敗）" } else { "" }; + steps.push(format!("{} {} → {}{failed}", c.name, cut(&c.args, 160), cut(result, 220))); + } + } + const KEEP: usize = 40; + if steps.len() > KEEP * 2 { + let gone = steps.len() - KEEP * 2; + steps.splice(KEEP..steps.len() - KEEP, [format!("…（中間省略 {gone} 步）…")]); + } + let ask = turn.first().map(|m| cut(&m.text, 800)).unwrap_or_default(); + let answer = turn.iter().rev().find(|m| m.role == crate::model::Role::Assistant && !m.text.trim().is_empty()).map(|m| cut(&m.text, 600)).unwrap_or_default(); + let mut s = format!("使用者的要求：\n{ask}\n\n每一步（工具 參數 → 結果）：\n"); + for (i, st) in steps.iter().enumerate() { + s.push_str(&format!("{}. {st}\n", i + 1)); + } + s.push_str(&format!("\n最後的回覆：\n{answer}\n\n已經有的技能：")); + if skills.is_empty() { + s.push_str("（沒有）"); + } + let mut room = 8000usize; + for k in skills { + let body = cut(&k.instructions, 2000.min(room)); + room = room.saturating_sub(body.chars().count()); + s.push_str(&format!("\n- {}：{}\n 做法：{body}", k.name, k.description)); + } + s +} + +/// The skill in the model's answer, when it chose to save one. +pub fn learned_skill(text: &str) -> Option { + let t = text.trim(); + let json = t.find('{').and_then(|a| t.rfind('}').filter(|b| *b >= a).map(|b| &t[a..=b]))?; + let v: serde_json::Value = serde_json::from_str(json).ok()?; + if v["save"].as_bool() != Some(true) { + return None; + } + let name: String = v["name"].as_str()?.trim().chars().take(40).collect(); + let instructions: String = v["instructions"].as_str()?.trim().chars().take(MAX_SKILL_CHARS).collect(); + if name.is_empty() || instructions.is_empty() { + return None; + } + Some(Skill { name, description: v["description"].as_str().unwrap_or("").trim().chars().take(120).collect(), instructions }) +} + +/// The line under the bot's reply that says it learned (or improved) a skill. +pub fn learned_note(name: &str, fresh: bool) -> String { + if fresh { + format!("\n\n（學會了新做法「{name}」，下次做這件事會照著做。）") + } else { + format!("\n\n（把做法「{name}」更新了，下次會照新的做。）") + } +} + pub fn prompt_section(memories: &[String], skills: &[Skill]) -> String { let mut s = String::new(); if !memories.is_empty() { @@ -74,7 +189,7 @@ pub fn prompt_section(memories: &[String], skills: &[Skill]) -> String { s.push_str("\n\n使用者說「記住…」「以後都…」時，用 remember 記下來；說「忘記…」時用 forget。"); } if !skills.is_empty() { - s.push_str("\n\n可用的技能（做法說明；需要時先用 use_skill 讀取完整內容再照做）：\n"); + s.push_str("\n\n你會的技能（以前做成功的做法）。開始做一件事之前先看這裡：有符合的，先用 use_skill 讀出完整做法再照著做，不要重新摸索：\n"); for k in skills { s.push_str(&format!("- {}：{}\n", k.name, k.description)); } @@ -102,6 +217,83 @@ mod tests { assert_eq!(m.len(), MAX_MEMORIES); assert_eq!(m[0], "事 5"); let p = prompt_section(&m[..1], &[Skill { name: "週報".into(), description: "整理每週進度".into(), instructions: "…".into() }]); - assert!(p.contains("1. 事 5") && p.contains("- 週報：整理每週進度")); + assert!(p.contains("1. 事 5") && p.contains("- 週報：整理每週進度") && p.contains("use_skill")); + } + + fn call(id: &str, name: &str, args: &str, failed: bool) -> crate::model::ToolCall { + crate::model::ToolCall { id: id.into(), name: name.into(), args: args.into(), status: if failed { crate::model::ToolStatus::Failed } else { crate::model::ToolStatus::Done } } + } + + fn result(id: &str, text: &str) -> crate::model::Message { + crate::model::Message { role: crate::model::Role::Tool, tool_call_id: id.into(), text: text.into(), ..Default::default() } + } + + /// A post with a picture: five real steps, one of them failed, the person corrects it midway. + fn post_turn() -> Vec { + use crate::model::Message; + let step = |c: crate::model::ToolCall| Message { tool_calls: vec![c], ..Message::assistant("", "b") }; + vec![ + Message::user("幫我發一篇 Threads，記得附圖"), + step(call("1", "browser_open", r#"{"url":"https://www.threads.com/"}"#, false)), + result("1", "已打開 https://www.threads.com/\n[13] 按鈕「Create」"), + step(call("2", "browser_click", r#"{"ref":"13"}"#, false)), + result("2", "已點 [13]。新出現的文字：New thread"), + Message { steer: true, ..Message::user("語氣要輕鬆一點") }, + step(call("3", "browser_type", r#"{"ref":"12","value":"今天天氣好"}"#, false)), + result("3", "已在 [12] 輸入。"), + step(call("4", "browser_upload", r#"{"path":"/home/bot/x.jpg"}"#, true)), + result("4", "這台電腦上沒有這個檔案"), + step(call("5", "browser_upload", r#"{"path":"/home/bot/bot-6/附件/x.jpg"}"#, false)), + result("5", "已把 /home/bot/bot-6/附件/x.jpg 放進上傳欄位。"), + step(call("6", "remember", r#"{"fact":"x"}"#, false)), + result("6", "記住了"), + Message::assistant("發好了：https://www.threads.com/@me/post/1", "b"), + ] + } + + #[test] + fn a_long_enough_job_is_looked_back_on() { + let turn = post_turn(); + assert!(worth_learning(&turn)); + // App steps (remember) don't count; four real steps aren't enough. + assert!(!worth_learning(&turn[..10])); + // A job that saved a skill itself is left alone. + let mut saved = turn.clone(); + saved[12].tool_calls[0].name = "save_skill".into(); + assert!(!worth_learning(&saved)); + } + + #[test] + fn the_look_back_has_every_step_and_the_skills() { + let skills = [Skill { name: "Threads 發文".into(), description: "在 Threads 發文".into(), instructions: "1. browser_open https://www.threads.com/".into() }]; + let b = learn_brief(&post_turn(), &skills); + assert!(b.starts_with("使用者的要求：\n幫我發一篇 Threads，記得附圖"), "{b}"); + assert!(b.contains("1. browser_open {\"url\":\"https://www.threads.com/\"} → 已打開 https://www.threads.com/ [13] 按鈕「Create」"), "{b}"); + assert!(b.contains("（使用者中途說：語氣要輕鬆一點）"), "{b}"); + assert!(b.contains("browser_upload {\"path\":\"/home/bot/x.jpg\"} → 這台電腦上沒有這個檔案（失敗）"), "{b}"); + assert!(b.contains("最後的回覆：\n發好了"), "{b}"); + assert!(b.contains("- Threads 發文：在 Threads 發文\n 做法：1. browser_open"), "{b}"); + // A very long job keeps its start and its end. + let mut long = vec![crate::model::Message::user("做很多事")]; + for i in 0..200 { + long.push(crate::model::Message { tool_calls: vec![call(&i.to_string(), "browser_click", "{}", false)], ..crate::model::Message::assistant("", "b") }); + } + let b = learn_brief(&long, &[]); + assert!(b.contains("（中間省略 120 步）") && b.contains("81. browser_click") && !b.contains("82. "), "{b}"); + } + + #[test] + fn a_learned_skill_is_read_from_the_answer() { + let k = learned_skill("好的\n{\"save\": true, \"name\": \" Threads 發文附圖 \", \"description\": \"在 Threads 發一篇附圖的文\", \"instructions\": \"1. browser_open https://www.threads.com/\"}").unwrap(); + assert_eq!((k.name.as_str(), k.description.as_str()), ("Threads 發文附圖", "在 Threads 發一篇附圖的文")); + assert!(learned_skill(r#"{"save": false}"#).is_none()); + assert!(learned_skill(r#"{"save": true, "name": "", "instructions": "x"}"#).is_none()); + assert!(learned_skill("不用存").is_none()); + assert!(learned_skill("} {").is_none()); + let mut list = vec![]; + assert!(upsert_skill(&mut list, k.clone())); + assert!(!upsert_skill(&mut list, Skill { instructions: "新的".into(), ..k })); + assert_eq!((list.len(), list[0].instructions.as_str()), (1, "新的")); + assert!(learned_note("Threads 發文附圖", true).contains("學會了新做法「Threads 發文附圖」")); } } diff --git a/src/providers.rs b/src/providers.rs index 8c96983..f331df6 100644 --- a/src/providers.rs +++ b/src/providers.rs @@ -199,6 +199,9 @@ pub enum StreamEvent { React(String), /// The bot sent a file into the chat (a file card under its reply). Attach(crate::attach::Attachment), + /// 自己進化: after the turn, the way that worked was kept as a skill (with the time of the + /// reply it learned from, which gets a line saying so). + Learned(crate::memory::Skill, i64), Done, Error(String), }