@nanhara/hara 0.33.0 → 0.48.0

package/dist/search/semindex.js

@@ -2,7 +2,7 @@
 // code-asset / repo / knowledge-base scale (hundreds–low-thousands of chunks); the optional zvec adapter is
 // the scale-up path later. Markdown/code stays the SSOT; this index is a derived, rebuildable, gitignored
 // artifact. The embedder is injected (see embed.ts) so the store + chunking are testable without a model.
-import { readFileSync, writeFileSync, mkdirSync, existsSync } from "node:fs";
+import { readFileSync, writeFileSync, mkdirSync, existsSync, statSync } from "node:fs";
 import { homedir } from "node:os";
 import { join, dirname } from "node:path";
 import { findProjectRoot } from "../context/agents-md.js";
@@ -15,14 +15,22 @@ export function indexPath(name, cwd) {
         return join(findProjectRoot(cwd), ".hara", "index", "repo.json");
     return join(homedir(), ".hara", "index", `${name}.json`);
 }
+function statMtime(p) {
+    try {
+        return statSync(p).mtimeMs;
+    }
+    catch {
+        return 0;
+    }
+}
 /** Split a file into chunks: Markdown by `#` headings, code by ~40-line windows. Heuristic, zero-dep —
- *  also the substrate embeddings reuse. */
-export function chunkText(text, file, source) {
+ *  also the substrate embeddings reuse. `mtime` (when given) is stamped on every chunk for incremental reuse. */
+export function chunkText(text, file, source, mtime) {
     const out = [];
     const push = (body, n) => {
         const t = body.trim();
         if (t.length >= 12)
-            out.push({ id: `${file}#${n}`, text: t.slice(0, 2000), file, source });
+            out.push({ id: `${file}#${n}`, text: t.slice(0, 2000), file, source, mtime });
     };
     if (/\.(md|mdx)$/i.test(file)) {
         const parts = text.split(/^(?=#{1,6}\s)/m);
@@ -49,23 +57,66 @@ function cosine(a, b) {
     }
     return na && nb ? dot / (Math.sqrt(na) * Math.sqrt(nb)) : 0;
 }
-/** Embed all chunks and write the index file. Returns the count written. */
+/** Build/refresh the index. **Incremental**: files whose mtime is unchanged since the last build keep their
+ *  existing vectors (no re-embed); only new/changed files are embedded, and deleted files drop out. A changed
+ *  embedding model forces a full rebuild (old vectors aren't comparable). Returns counts. */
 export async function buildIndex(name, chunks, embed, cwd, model = "embed") {
+    const p = indexPath(name, cwd);
+    // Load the previous index → reuse vectors for unchanged files.
+    const prevByFile = new Map();
+    let prevModel = "";
+    if (existsSync(p)) {
+        try {
+            const old = JSON.parse(readFileSync(p, "utf8"));
+            prevModel = old.model;
+            for (const it of old.items ?? []) {
+                const arr = prevByFile.get(it.file);
+                if (arr)
+                    arr.push(it);
+                else
+                    prevByFile.set(it.file, [it]);
+            }
+        }
+        catch {
+            /* corrupt index → full rebuild */
+        }
+    }
+    const sameModel = prevModel === model;
+    const byFile = new Map();
+    for (const c of chunks) {
+        const arr = byFile.get(c.file);
+        if (arr)
+            arr.push(c);
+        else
+            byFile.set(c.file, [c]);
+    }
     const items = [];
+    const toEmbed = [];
+    let reused = 0;
+    for (const [file, fchunks] of byFile) {
+        const mtime = fchunks[0].mtime ?? 0;
+        const prev = prevByFile.get(file);
+        if (sameModel && prev?.length && mtime > 0 && prev.every((it) => it.mtime === mtime)) {
+            items.push(...prev); // file unchanged → keep its vectors
+            reused += prev.length;
+        }
+        else {
+            toEmbed.push(...fchunks);
+        }
+    }
     const B = 64;
-    for (let i = 0; i < chunks.length; i += B) {
-        const batch = chunks.slice(i, i + B);
+    for (let i = 0; i < toEmbed.length; i += B) {
+        const batch = toEmbed.slice(i, i + B);
         const vecs = await embed(batch.map((c) => c.text));
         batch.forEach((c, j) => vecs[j] && items.push({ ...c, vec: vecs[j] }));
     }
-    const p = indexPath(name, cwd);
     const dir = dirname(p);
     mkdirSync(dir, { recursive: true });
     // The index is derived + rebuildable (and may embed file contents) — never let it be committed.
     if (!existsSync(join(dir, ".gitignore")))
         writeFileSync(join(dir, ".gitignore"), "*\n", "utf8");
     writeFileSync(p, JSON.stringify({ model, items }), "utf8");
-    return items.length;
+    return { total: items.length, embedded: toEmbed.length, reused };
 }
 export function indexExists(name, cwd) {
     return existsSync(indexPath(name, cwd));
@@ -91,7 +142,7 @@ export function collectDirChunks(dir, source) {
         }
         if (isProbablyBinary(buf))
             continue;
-        chunks.push(...chunkText(buf.toString("utf8"), abs, source));
+        chunks.push(...chunkText(buf.toString("utf8"), abs, source, statMtime(abs)));
     }
     return chunks;
 }
@@ -113,7 +164,7 @@ export function collectRepoChunks(root) {
         }
         if (isProbablyBinary(buf))
             continue;
-        chunks.push(...chunkText(buf.toString("utf8"), rel, "repo"));
+        chunks.push(...chunkText(buf.toString("utf8"), rel, "repo", statMtime(abs)));
     }
     return chunks;
 }

package/dist/session/store.js

@@ -56,6 +56,20 @@ export function titleFrom(history) {
     const firstUser = history.find((h) => h.role === "user");
     return deriveTitle(firstUser && firstUser.role === "user" ? firstUser.content : "");
 }
+/** Normalize a phrase to an ASCII kebab-case slug (lowercase, a–z0–9 + single hyphens, capped). Non-ASCII
+ *  is dropped — used to clean a model-generated English session name. Returns "" if nothing ASCII remains. */
+export function slugify(text, max = 40) {
+    return text
+        .trim()
+        .toLowerCase()
+        .replace(/[^a-z0-9\s-]/g, "")
+        .trim()
+        .replace(/\s+/g, "-")
+        .replace(/-+/g, "-")
+        .replace(/^-+/, "")
+        .slice(0, max)
+        .replace(/-+$/, "");
+}
 export function saveSession(meta, history) {
     meta.updatedAt = new Date().toISOString();
     const data = { meta, history };

package/dist/tools/computer.js

@@ -11,7 +11,7 @@ import { join } from "node:path";
 import { registerTool } from "./registry.js";
 import { loadConfig } from "../config.js";
 const RANK = { off: 0, read: 1, click: 2, full: 3 };
-const ACTION_MIN = { screenshot: "read", move: "click", click: "click", type: "full", key: "full" };
+const ACTION_MIN = { screenshot: "read", find: "read", activate: "click", move: "click", click: "click", type: "full", key: "full" };
 // dangerous combos refused even at full tier (quit / close / delete / task-switch-kill)
 const KEY_BLOCK = /(?:\b(cmd|command|ctrl|control|alt|option|win|super|meta)\b.*\+.*\b(q|w|delete|del|f4|escape|esc)\b)|ctrl\+alt\+(?:delete|del|backspace)/i;
 /** Whether the configured tier permits the action. Exported for tests. */
@@ -22,6 +22,25 @@ export function actionAllowed(tier, action) {
 export function keyIsBlocked(keys) {
     return KEY_BLOCK.test(keys);
 }
+// Circuit breaker (learned from codex): bound consecutive screen-control failures so the agent can't loop
+// forever on a broken setup. Reset on any success; after FAIL_LIMIT in a row, return a clear stop + how to fix.
+const FAIL_LIMIT = 3;
+let consecFails = 0;
+export function resetComputerFails() {
+    consecFails = 0;
+}
+function ok(msg) {
+    consecFails = 0;
+    return msg;
+}
+function fail(msg) {
+    consecFails += 1;
+    if (consecFails >= FAIL_LIMIT) {
+        consecFails = 0;
+        return `⛔ Stopping screen control — ${FAIL_LIMIT} actions failed in a row (last: ${msg}). Most likely a missing macOS permission (Accessibility for click/type, Screen Recording for screenshots) or the target app isn't reachable. Fix that, then ask me to try again — I won't keep retrying blindly.`;
+    }
+    return `Failed: ${msg}  [${consecFails}/${FAIL_LIMIT} before I stop]`;
+}
 function run(cmd, args) {
     try {
         const r = spawnSync(cmd, args, { encoding: "utf8", timeout: 15000 });
@@ -35,6 +54,23 @@ function has(cmd) {
     return (process.platform === "win32" ? run("where", [cmd]) : run("which", [cmd])).ok;
 }
 const ps = (script) => run("powershell", ["-NoProfile", "-Command", script]);
+/** Put text on the OS clipboard (so `type` can paste it — IME-safe + Unicode-safe, unlike keystroke injection). */
+function setClipboard(text) {
+    try {
+        if (process.platform === "darwin")
+            return spawnSync("pbcopy", [], { input: text, timeout: 5000 }).status === 0;
+        if (process.platform === "win32")
+            return spawnSync("clip", [], { input: text, timeout: 5000 }).status === 0;
+        if (has("wl-copy"))
+            return spawnSync("wl-copy", [], { input: text, timeout: 5000 }).status === 0;
+        if (has("xclip"))
+            return spawnSync("xclip", ["-selection", "clipboard"], { input: text, timeout: 5000 }).status === 0;
+    }
+    catch {
+        /* fall through */
+    }
+    return false;
+}
 let seq = 0;
 function tmpShot() {
     seq += 1;
@@ -73,6 +109,49 @@ function screenshot() {
     }
     return { path: out };
 }
+/** Bring an app to the foreground so screenshots/clicks land on IT, not the terminal hara runs in. */
+function activateApp(app) {
+    if (process.platform === "darwin") {
+        // `open -a` reliably launches+foregrounds; `osascript … activate` often leaves another window on top.
+        const r = run("open", ["-a", app]);
+        return { ok: r.ok, msg: r.ok ? `activated ${app}` : r.out || `couldn't activate ${app}` };
+    }
+    if (process.platform === "win32") {
+        const r = ps(`(New-Object -ComObject WScript.Shell).AppActivate(${JSON.stringify(app)})`);
+        return { ok: r.ok, msg: r.ok ? `activated ${app}` : r.out || `couldn't activate ${app}` };
+    }
+    if (process.platform === "linux") {
+        const r = has("wmctrl") ? run("wmctrl", ["-a", app]) : run("xdotool", ["search", "--name", app, "windowactivate"]);
+        return { ok: r.ok, msg: r.ok ? `activated ${app}` : r.out || `couldn't activate ${app} (need wmctrl/xdotool)` };
+    }
+    return { ok: false, msg: `activate unsupported on ${process.platform}` };
+}
+/** Logical screen size in the coordinate space the click backends use (points on mac, pixels on win/linux).
+ *  Grounding returns 0..1 fractions, so click = fraction × this. null if undetectable. */
+function screenSize() {
+    try {
+        if (process.platform === "darwin") {
+            const r = run("osascript", ["-e", 'tell application "Finder" to get bounds of window of desktop']);
+            const n = r.out.match(/-?\d+/g);
+            if (n && n.length >= 4)
+                return { w: Number(n[2]), h: Number(n[3]) };
+        }
+        else if (process.platform === "linux") {
+            const [w, h] = run("xdotool", ["getdisplaygeometry"]).out.trim().split(/\s+/).map(Number);
+            if (w && h)
+                return { w, h };
+        }
+        else if (process.platform === "win32") {
+            const [w, h] = ps('Add-Type -AssemblyName System.Windows.Forms; $b=[System.Windows.Forms.Screen]::PrimaryScreen.Bounds; "$($b.Width) $($b.Height)"').out.trim().split(/\s+/).map(Number);
+            if (w && h)
+                return { w, h };
+        }
+    }
+    catch {
+        /* fall through */
+    }
+    return null;
+}
 /** Name of the frontmost application/window (for the allowlist check). "" if undetectable. */
 function frontmostApp() {
     if (process.platform === "darwin") {
@@ -128,21 +207,41 @@ using System;using System.Runtime.InteropServices;public class Ms{[DllImport("us
         const text = String(input.text ?? "");
         if (!text)
             return { ok: false, msg: "type needs text" };
+        // IME-safe path: set the clipboard and paste. Keystroke injection (below) is intercepted/garbled by a
+        // CJK input method and can't enter Chinese/emoji reliably; pasting is immune and Unicode-safe.
+        if (setClipboard(text)) {
+            if (mac && has("cliclick")) {
+                const r = run("cliclick", ["kd:cmd", "t:v", "ku:cmd"]); // Cmd+V
+                if (r.ok)
+                    return { ok: true, msg: `pasted ${text.length} chars` };
+            }
+            else if (lin && has("xdotool")) {
+                const r = run("xdotool", ["key", "ctrl+v"]);
+                if (r.ok)
+                    return { ok: true, msg: `pasted ${text.length} chars` };
+            }
+            else if (win) {
+                const r = ps("Add-Type -AssemblyName System.Windows.Forms;[System.Windows.Forms.SendKeys]::SendWait('^v')");
+                if (r.ok)
+                    return { ok: true, msg: `pasted ${text.length} chars` };
+            }
+        }
+        // Fallback: keystroke injection (fine for ASCII when no IME is active).
         if (mac) {
             if (!has("cliclick"))
                 return { ok: false, msg: "cliclick not found — install with `brew install cliclick`" };
             const r = run("cliclick", [`t:${text}`]);
-            return { ok: r.ok, msg: r.ok ? `typed ${text.length} chars` : r.out };
+            return { ok: r.ok, msg: r.ok ? `typed ${text.length} chars (keystroke)` : r.out };
         }
         if (lin) {
             if (!has("xdotool"))
                 return { ok: false, msg: "xdotool not found" };
             const r = run("xdotool", ["type", "--clearmodifiers", text]);
-            return { ok: r.ok, msg: r.ok ? `typed ${text.length} chars` : r.out };
+            return { ok: r.ok, msg: r.ok ? `typed ${text.length} chars (keystroke)` : r.out };
         }
         if (win) {
             const r = ps(`Add-Type -AssemblyName System.Windows.Forms;[System.Windows.Forms.SendKeys]::SendWait(${JSON.stringify(text)})`);
-            return { ok: r.ok, msg: r.ok ? `typed ${text.length} chars` : r.out };
+            return { ok: r.ok, msg: r.ok ? `typed ${text.length} chars (keystroke)` : r.out };
         }
     }
     if (action === "key") {
@@ -182,17 +281,23 @@ export function computerBackends() {
 }
 registerTool({
     name: "computer",
-    description: "Control the screen to operate desktop software (not just the browser): take a screenshot, then " +
-        "click/move/type/press keys at coordinates. Workflow: screenshot → read what's on screen → act. " +
-        "Opt-in and permission-gated (tier + per-app allowlist).",
+    description: "Control the screen to operate desktop software (not just the browser). ALWAYS `activate` the target app " +
+        "FIRST (e.g. activate WeChat) — otherwise screenshots/clicks hit the terminal hara runs in, not the app. " +
+        "Then prefer grounding over guessing pixels: pass `target` (e.g. 'the Send button') to click/move and it's " +
+        "located by a vision model; or `find` to just get coordinates. Workflow: activate → screenshot → click a " +
+        "target → re-screenshot to verify. When typing, type the ACTUAL text — never placeholders. Opt-in and " +
+        "permission-gated (tier + per-app allowlist).",
     input_schema: {
         type: "object",
         properties: {
-            action: { type: "string", enum: ["screenshot", "click", "move", "type", "key"] },
-            x: { type: "number", description: "x pixel (click/move)" },
-            y: { type: "number", description: "y pixel (click/move)" },
+            action: { type: "string", enum: ["screenshot", "activate", "find", "click", "move", "type", "key"] },
+            app: { type: "string", description: "app to bring to the foreground (activate) — e.g. 'WeChat'. Do this BEFORE screenshot/click so they hit the app, not the terminal." },
+            target: { type: "string", description: "describe a UI element to locate (find) or click/move to — e.g. 'the Send button'. Preferred over x,y." },
+            x: { type: "number", description: "x pixel (click/move; or use `target`)" },
+            y: { type: "number", description: "y pixel (click/move; or use `target`)" },
             text: { type: "string", description: "text to type (type)" },
             keys: { type: "string", description: "key or combo, e.g. 'return', 'cmd+c' (key)" },
+            focus: { type: "string", description: "screenshot only: what to look for — focuses the read" },
         },
         required: ["action"],
     },
@@ -205,7 +310,17 @@ registerTool({
         const action = String(input.action ?? "");
         if (!actionAllowed(tier, action))
             return `'${action}' needs a higher tier (current computerUse=${tier}). Raise it with \`hara config set computerUse …\`.`;
-        if (action !== "screenshot") {
+        // Bring the target app to the foreground first — without this, clicks land on the terminal hara runs in.
+        if (action === "activate") {
+            const app = String(input.app ?? input.target ?? "");
+            if (!app)
+                return "activate needs an `app` name (e.g. 'WeChat').";
+            if (!cfg.computerApps.some((a) => app.toLowerCase().includes(a.toLowerCase()) || a.toLowerCase().includes(app.toLowerCase())))
+                return `Refused: "${app}" isn't in your allowlist (${cfg.computerApps.join(", ") || "empty"}). Add it: \`hara config set computerApps "${app}"\`.`;
+            const r = activateApp(app);
+            return r.ok ? ok(`✓ ${r.msg} — now screenshot/find/click to act on it`) : fail(r.msg);
+        }
+        if (action !== "screenshot" && action !== "find") {
             // per-app allowlist: only act when an allowlisted app is frontmost (the key guard against wrong-window clicks)
             if (!cfg.computerApps.length)
                 return "No apps allowlisted — set `hara config set computerApps \"App Name, …\"` before clicking/typing.";
@@ -217,20 +332,45 @@ registerTool({
         if (action === "screenshot") {
             const s = screenshot();
             if (s.error)
-                return `Screenshot failed: ${s.error}`;
+                return fail(`screenshot — ${s.error}`);
             if (ctx.describeImage) {
                 try {
-                    const desc = await ctx.describeImage(s.path);
+                    const desc = await ctx.describeImage(s.path, input.focus ? String(input.focus) : undefined);
                     if (desc)
-                        return `Screenshot (read via vision):\n${desc}`;
+                        return ok(`Screenshot (read via vision):\n${desc}`);
                 }
                 catch {
                     /* fall through to path */
                 }
             }
-            return `Screenshot saved to ${s.path}. Configure a vision model so I can read it: \`hara config set visionModel <model>\`.`;
+            return ok(`Screenshot saved to ${s.path}. Configure a vision model so I can read it: \`hara config set visionModel <model>\`.`);
+        }
+        // Grounding: locate a described element and turn it into screen coordinates (more reliable than guessing
+        // pixels from a text description). Used for `find`, and for click/move when given a `target` and no x,y.
+        const needsLocate = action === "find" || ((action === "click" || action === "move") && input.target != null && (input.x == null || input.y == null));
+        if (needsLocate) {
+            const target = String(input.target ?? "");
+            if (!target)
+                return action === "find" ? "find needs a `target` (what to locate)." : "click/move needs `x,y` or a `target`.";
+            if (!ctx.locate)
+                return "Grounding needs a vision model that can see images — set one: `hara config set visionModel <model>`.";
+            const s = screenshot();
+            if (s.error)
+                return fail(`screenshot — ${s.error}`);
+            const loc = await ctx.locate(s.path, target);
+            if (!loc)
+                return fail(`couldn't locate "${target}" on screen — try a screenshot first, or rephrase the target`);
+            const size = screenSize();
+            if (!size)
+                return fail(`located "${target}" but couldn't read the screen size to convert coordinates`);
+            const gx = Math.round(loc.x * size.w);
+            const gy = Math.round(loc.y * size.h);
+            if (action === "find")
+                return ok(`"${target}" is at ~${gx},${gy} (${Math.round(loc.x * 100)}% across, ${Math.round(loc.y * 100)}% down).`);
+            input.x = gx;
+            input.y = gy;
         }
         const r = pointerOrKeyboard(action, input);
-        return r.ok ? `✓ ${r.msg}` : `Failed: ${r.msg}`;
+        return r.ok ? ok(`✓ ${r.msg}${needsLocate ? ` (located "${input.target}")` : ""}`) : fail(r.msg);
     },
 });

package/dist/tui/App.js

@@ -13,6 +13,7 @@ import { InputBox } from "./InputBox.js";
 import { activity } from "../activity.js";
 import { ctxPctFor } from "../statusbar.js";
 import { accent } from "./theme.js";
+import { renderMarkdown } from "../md.js";
 let _id = 0;
 const nid = () => ++_id;
 const stripAnsi = (s) => s.replace(/\x1b\[[0-9;]*m/g, "");
@@ -21,7 +22,7 @@ function Block({ item, open }) {
         case "user":
             return (_jsxs(Box, { marginTop: 1, children: [_jsx(Text, { color: "cyan", children: "\u203A " }), _jsx(Text, { children: item.text })] }));
         case "assistant":
-            return _jsx(Text, { children: item.text });
+            return _jsx(Text, { children: renderMarkdown(item.text) }); // headers/bold/inline-code/bullets + verbatim fences
         case "reasoning": {
             // fixed-height window: show the last 5 lines while thinking; ctrl-r toggles the full text.
             const lines = item.text.replace(/\n+$/, "").split("\n");
@@ -71,6 +72,9 @@ export function App({ initialStatus, model, cwd, header, onSubmit, cycleApproval
     const [promptSel, setPromptSel] = useState(0);
     const [reasoningOpen, setReasoningOpen] = useState(false);
     const ctrlRef = useRef(null);
+    const queueRef = useRef([]); // type-ahead: FIFO of messages entered while working
+    const [pool, setPool] = useState([]); // type-ahead pool: queued message lines, shown above the input
+    const drainingRef = useRef(false); // idempotency guard so the drain effect can't double-send one item
     const currentRef = useRef([]);
     currentRef.current = current;
     const statusRef = useRef(status);
@@ -90,8 +94,14 @@ export function App({ initialStatus, model, cwd, header, onSubmit, cycleApproval
     }, []);
     const handleSubmit = useCallback(async (line, images) => {
         const t = line.trim();
-        if ((!t && !images?.length) || working || prompt)
-            return; // allow image-only turns
+        if ((!t && !images?.length) || prompt)
+            return; // nothing to send, or a choice is pending
+        if (working) {
+            // type-ahead: hold the message in the pool; all pooled messages are sent together when the turn ends
+            queueRef.current.push({ line, images });
+            setPool(queueRef.current.map((q) => q.line.trim() || "🖼 (image)"));
+            return;
+        }
         setHistory((h) => [...h, { id: nid(), kind: "user", text: t }]); // t already carries any [Image #N] tokens
         const ctrl = new AbortController();
         ctrlRef.current = ctrl;
@@ -130,6 +140,21 @@ export function App({ initialStatus, model, cwd, header, onSubmit, cycleApproval
         setWorking(false);
         ctrlRef.current = null;
     }, [working, prompt, onSubmit, pushCurrent, model, exit]);
+    // Drain the type-ahead pool: when the turn finishes (working → false) and nothing awaits a choice, COALESCE
+    // every pooled message into ONE turn and send it — additions/clarifications go to the agent together, in order.
+    useEffect(() => {
+        if (working || prompt || drainingRef.current || !queueRef.current.length)
+            return;
+        drainingRef.current = true;
+        const batch = queueRef.current;
+        queueRef.current = [];
+        setPool([]);
+        const line = batch.map((b) => b.line).join("\n\n");
+        const images = batch.flatMap((b) => b.images ?? []);
+        void Promise.resolve(handleSubmit(line, images.length ? images : undefined)).finally(() => {
+            drainingRef.current = false;
+        });
+    }, [working, prompt, handleSubmit]);
     useInput((input, key) => {
         if (prompt) {
             const opts = prompt.options;
@@ -145,6 +170,10 @@ export function App({ initialStatus, model, cwd, header, onSubmit, cycleApproval
                 prompt.resolve(opts[opts.length - 1].value); // last option = cancel/no
                 setPrompt(null);
             }
+            else if (/^[1-9]$/.test(input) && Number(input) <= opts.length) {
+                prompt.resolve(opts[Number(input) - 1].value); // type a number to pick directly
+                setPrompt(null);
+            }
             else if (input) {
                 const hit = opts.find((o) => o.key && o.key === input.toLowerCase());
                 if (hit) {
@@ -156,10 +185,16 @@ export function App({ initialStatus, model, cwd, header, onSubmit, cycleApproval
         }
         if (key.ctrl && input === "r")
             return setReasoningOpen((x) => !x);
-        if (key.escape && working)
+        if (key.escape && working) {
+            // Esc = stop everything: abort the turn AND drop any type-ahead (a stopped turn shouldn't fire queued msgs)
+            if (queueRef.current.length) {
+                queueRef.current = [];
+                setPool([]);
+            }
             ctrlRef.current?.abort();
+        }
         else if (key.tab && key.shift && cycleApproval)
             setStatus((s) => ({ ...s, approval: cycleApproval(s.approval) }));
     });
-    return (_jsxs(Box, { flexDirection: "column", children: [_jsx(Static, { items: header ? [{ id: -1, kind: "notice", text: "" }, ...history] : history, children: (item) => (item.id === -1 ? _jsx(HeaderCard, { ...header }, "hdr") : _jsx(Block, { item: item }, item.id)) }), current.map((item) => (_jsx(Block, { item: item, open: reasoningOpen }, item.id))), working && !prompt && _jsx(Working, {}), prompt && (_jsxs(Box, { flexDirection: "column", marginTop: 1, children: [_jsx(Text, { color: "yellow", children: `  ${stripAnsi(prompt.title)}` }), prompt.options.map((o, i) => (_jsx(Text, { color: i === promptSel ? "cyan" : undefined, bold: i === promptSel, children: (i === promptSel ? " ❯ " : "   ") + o.label }, i)))] })), _jsx(InputBox, { status: status, cwd: cwd, isActive: !working && !prompt, onSubmit: handleSubmit, onClipboardImage: onClipboardImage })] }));
+    return (_jsxs(Box, { flexDirection: "column", children: [_jsx(Static, { items: header ? [{ id: -1, kind: "notice", text: "" }, ...history] : history, children: (item) => (item.id === -1 ? _jsx(HeaderCard, { ...header }, "hdr") : _jsx(Block, { item: item }, item.id)) }), current.map((item) => (_jsx(Block, { item: item, open: reasoningOpen }, item.id))), working && !prompt && _jsx(Working, {}), prompt && (_jsxs(Box, { flexDirection: "column", marginTop: 1, children: [_jsx(Text, { color: "yellow", children: `  ${stripAnsi(prompt.title)}` }), prompt.options.map((o, i) => (_jsx(Text, { color: i === promptSel ? "cyan" : undefined, bold: i === promptSel, children: (i === promptSel ? " ❯ " : "   ") + `${i + 1}. ` + o.label }, i))), _jsx(Text, { dimColor: true, children: `   ↑↓ or 1–${prompt.options.length} to choose · Enter · Esc cancels` })] })), pool.length > 0 && !prompt && (_jsx(Box, { flexDirection: "column", children: pool.map((l, i) => (_jsx(Text, { color: accent(), children: `  › ${l.length > 72 ? l.slice(0, 72) + "…" : l}` }, i))) })), _jsx(InputBox, { status: status, cwd: cwd, isActive: !prompt, working: working, queued: pool.length, onSubmit: handleSubmit, onClipboardImage: onClipboardImage })] }));
 }

package/dist/tui/InputBox.js

@@ -84,7 +84,7 @@ function InputLine({ value, cursor }) {
     return _jsx(Text, { children: nodes });
 }
 /** Top border (session) + prompt line + bottom border (usage) + ModeBar, with an @path popup. */
-export function InputBox({ status, cwd, width, onSubmit, onClipboardImage, isActive = true, placeholder = "Type a task · /help · @file · Ctrl+V paste image · shift+tab mode · Esc interrupts", }) {
+export function InputBox({ status, cwd, width, onSubmit, onClipboardImage, isActive = true, working = false, queued = 0, placeholder = "Type a task · /help · @file · Ctrl+V paste image · shift+tab mode · Esc interrupts", }) {
     const { stdout } = useStdout();
     const w = width ?? stdout?.columns ?? 80;
     const [value, setValue] = useState("");
@@ -204,5 +204,5 @@ export function InputBox({ status, cwd, width, onSubmit, onClipboardImage, isAct
             set(value.slice(0, cursor) + input + value.slice(cursor), cursor + input.length);
         }
     }, { isActive });
-    return (_jsxs(Box, { flexDirection: "column", children: [_jsx(TopBorder, { name: status.sessionName || "session", width: w }), _jsxs(Box, { children: [_jsx(Text, { color: "cyan", children: "› " }), value.length === 0 ? (_jsxs(Text, { children: [_jsx(Text, { inverse: true, children: " " }), _jsx(Text, { dimColor: true, children: placeholder })] })) : (_jsx(InputLine, { value: value, cursor: cursor }))] }), _jsx(BottomBorder, { s: status, width: w }), popupOpen ? _jsx(MentionPopup, { items: candidates, selected: selIdx, query: mention.query }) : null, _jsx(ModeBar, { approval: status.approval })] }));
+    return (_jsxs(Box, { flexDirection: "column", children: [_jsx(TopBorder, { name: status.sessionName || "session", width: w }), _jsxs(Box, { children: [_jsx(Text, { color: "cyan", children: "› " }), value.length === 0 ? (_jsxs(Text, { children: [_jsx(Text, { inverse: true, children: " " }), _jsx(Text, { dimColor: true, children: placeholder })] })) : (_jsx(InputLine, { value: value, cursor: cursor }))] }), _jsx(BottomBorder, { s: status, width: w }), working ? _jsx(Text, { dimColor: true, children: `  ⌨ working — Enter queues your message${queued ? ` · ${queued} queued` : ""} · Esc interrupts` }) : null, popupOpen ? _jsx(MentionPopup, { items: candidates, selected: selIdx, query: mention.query }) : null, _jsx(ModeBar, { approval: status.approval })] }));
 }

package/dist/vision.js

@@ -65,12 +65,61 @@ export const DESCRIBE_SYSTEM = [
     "4. Quote any error or warning messages exactly.",
     "5. Be thorough and factual; do not speculate beyond what is visible.",
 ].join("\n");
+// Screenshot variant — tuned for driving the desktop (RPA) rather than transcription. A text-only main
+// model can't see, so it needs *actionable* output: where things are, so it can issue clicks.
+export const SCREENSHOT_SYSTEM = [
+    "You are the eyes of an assistant operating this computer; it cannot see the screen and acts only on your",
+    "words. Describe the screenshot so it can ACT. Prioritise, in order:",
+    "1. INTERACTIVE elements — buttons, links, text fields, checkboxes, menus, tabs, icons — each with its",
+    "   visible label and an approximate location: a region (e.g. top-right) AND a rough pixel x,y if you can.",
+    "2. The currently focused/active element or selection, and any open dialog/modal/popup.",
+    "3. Errors, warnings, and key visible text/headings — quote them exactly.",
+    "4. One line on what app/screen this appears to be.",
+    "Positions guide clicks, so always estimate them. Be concise and factual; never invent elements.",
+].join("\n");
+// Grounding — ask a vision model WHERE a UI element is (for accurate RPA clicking), as resolution-independent
+// fractions so it works regardless of Retina/DPI scaling.
+export const LOCATE_SYSTEM = [
+    "You are given a screenshot. The user names ONE UI element (button, field, icon, menu item, link).",
+    "Return ONLY its CENTER as JSON: {\"x\": <0-1000>, \"y\": <0-1000>}, where x is the position as per-mille of",
+    "the image WIDTH (0=left, 1000=right) and y as per-mille of the HEIGHT (0=top, 1000=bottom).",
+    "If the element is not visible, return {\"x\": -1, \"y\": -1}. Output ONLY the JSON, nothing else.",
+].join("\n");
+/** Parse a grounding reply → {x,y} as 0..1 fractions (accepts per-mille / percent / fraction), or null. */
+export function parseLocate(text) {
+    const m = text.match(/"x"\s*:\s*(-?\d+(?:\.\d+)?)[\s,}]+.*?"y"\s*:\s*(-?\d+(?:\.\d+)?)/s) || text.match(/(-?\d+(?:\.\d+)?)\s*[,\s]\s*(-?\d+(?:\.\d+)?)/);
+    if (!m)
+        return null;
+    let x = Number(m[1]);
+    let y = Number(m[2]);
+    if (x < 0 || y < 0 || Number.isNaN(x) || Number.isNaN(y))
+        return null; // not found / unparseable
+    const norm = (v) => (v > 100 ? v / 1000 : v > 1.5 ? v / 100 : v); // per-mille | percent | fraction → 0..1
+    x = Math.min(1, Math.max(0, norm(x)));
+    y = Math.min(1, Math.max(0, norm(y)));
+    return { x, y };
+}
+/** Send a screenshot to a (grounding-capable) vision model and get the target's center as 0..1 fractions. */
+export async function locateImage(provider, image, target, opts = {}) {
+    const r = await provider.turn({
+        system: LOCATE_SYSTEM,
+        history: [{ role: "user", content: `Locate this element: ${target}`, images: [image] }],
+        tools: [],
+        onText: () => { },
+        signal: opts.signal,
+    });
+    if (r.stop === "error")
+        return null;
+    return parseLocate(r.text);
+}
 const PROMPT = "Describe the attached image(s) per your instructions.";
-/** Send images to the vision provider and return its textual description. Throws on a provider error. */
+/** Send images to the vision provider and return its textual description. Throws on a provider error.
+ *  `system` overrides the default prompt (e.g. SCREENSHOT_SYSTEM); `hint` focuses it on a specific goal. */
 export async function describeImages(provider, images, opts = {}) {
+    const content = opts.hint ? `${PROMPT}\nFocus especially on: ${opts.hint}` : PROMPT;
     const r = await provider.turn({
-        system: DESCRIBE_SYSTEM,
-        history: [{ role: "user", content: PROMPT, images }],
+        system: opts.system ?? DESCRIBE_SYSTEM,
+        history: [{ role: "user", content, images }],
         tools: [],
         onText: () => { },
         signal: opts.signal,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nanhara/hara",
-  "version": "0.33.0",
+  "version": "0.48.0",
   "description": "hara — a coding agent CLI that runs like an engineering org.",
   "bin": {
     "hara": "dist/index.js"
@@ -11,7 +11,8 @@
     "README.md",
     "CHANGELOG.md",
     "LICENSE",
-    "CLA.md"
+    "CLA.md",
+    "plugins"
   ],
   "keywords": [
     "ai",

package/plugins/browser/.hara-plugin/plugin.json

@@ -0,0 +1,9 @@
+{
+  "name": "browser",
+  "version": "0.1.0",
+  "description": "Reliable web automation for hara via the Playwright MCP — acts on the DOM/accessibility tree (selectors, auto-wait), not pixels. navigate / click / type / fill / snapshot.",
+  "skills": ["skills"],
+  "mcpServers": {
+    "browser": { "command": "npx", "args": ["-y", "@playwright/mcp@latest"] }
+  }
+}

package/plugins/browser/skills/web/SKILL.md

@@ -0,0 +1,27 @@
+---
+name: web-automation
+description: Operate web pages reliably — navigate, click, fill forms, log in, extract — via the Playwright MCP. Acts on the DOM/accessibility tree by selector/role (deterministic, auto-waiting), NOT screenshots or pixel coordinates. Far more reliable than desktop screen control.
+when_to_use: when the user wants to do anything on a website — open a page, click, fill/submit a form, log in, scrape data, automate a web flow.
+---
+
+# Web automation (Playwright MCP)
+
+Reliable browser tools are available as `mcp__browser__*` (navigate, snapshot, click, type, fill_form,
+select_option, evaluate, …). They act on the page's **accessibility tree by element ref/role/text** — not
+screenshots or pixel coordinates — so they're deterministic and auto-wait for elements. This is the reliable
+counterpart to the fragile desktop `computer` tool: prefer it for anything on the web.
+
+## Workflow
+1. `browser_navigate` to the URL.
+2. `browser_snapshot` — read the accessibility tree (elements + their `ref`s). This is your "eyes": use the
+   refs to act precisely. Prefer it over a screenshot.
+3. Act by ref/role/text: `browser_click`, `browser_type`, `browser_fill_form`, `browser_select_option`.
+4. `browser_snapshot` again to verify before the next step.
+
+## Notes
+- First run downloads a browser once: `npx playwright install chromium`.
+- The Playwright MCP uses its **own** browser (no logins). For tasks needing your **real logged-in Chrome**, use
+  `chrome-devtools-mcp` instead (drives your actual Chrome via CDP) — swap the mcpServers command to
+  `npx chrome-devtools-mcp@latest`. (This is what openclaw/cc-haha use.)
+- **Confirm before irreversible actions** — purchases, posting, sending messages, deleting. Verify the page/state
+  with a snapshot first.

package/plugins/chrome/.hara-plugin/plugin.json

@@ -0,0 +1,9 @@
+{
+  "name": "chrome",
+  "version": "0.1.0",
+  "description": "Drive a real, persistent-login Chrome from hara via chrome-devtools-mcp (CDP) — for web tasks on sites you're already signed into (logins persist across runs). Alternative to the `browser` plugin's isolated Playwright browser — enable one, not both.",
+  "skills": ["skills"],
+  "mcpServers": {
+    "chrome": { "command": "npx", "args": ["-y", "chrome-devtools-mcp@latest"] }
+  }
+}