autodev-cli 1.4.67 → 1.4.69

package/out/providers/grokTuiProvider.js

@@ -1,25 +1,40 @@
 "use strict";
 // ---------------------------------------------------------------------------
-// grokProvider -- Grok tasks via the `grok` headless CLI. Backs BOTH provider
-// variants from one shared runner:
+// grokProvider -- Grok tasks via the `grok` CLI. Backs BOTH provider variants:
 //
-//   grok-cli  (stateless):  fresh process each task, no session flags — no
-//                           context accumulation across tasks.
-//   grok-tui  (persistent): resumes ONE session per workspace so context
-//                           carries across tasks, like an interactive session.
-//                           First task: `--session-id <uuid>` (created + saved
-//                           to session-state.json); later tasks: `--resume <id>`.
+//   grok-cli  (stateless):  fresh HEADLESS process each task (--prompt-file),
+//                           no context accumulation across tasks.
+//   grok-tui  (persistent): ONE long-lived interactive `grok` process kept alive
+//                           inside a per-workspace tmux session (a real PTY), so
+//                           grok's in-process context accumulates across tasks —
+//                           exactly like a human-driven interactive session.
 //
-// Command:
-//   grok -m <model> --always-approve --cwd <root>
-//        [--session-id <uuid> | --resume <uuid>]      (grok-tui only)
-//        --prompt-file <file> --output-format streaming-json
+// -------------------------------------------------------------------------
+// Why tmux for grok-tui?
+// -------------------------------------------------------------------------
+// grok is an interactive terminal UI. The OLD grok-tui ran it HEADLESS with
+// piped, non-TTY stdio (`--prompt-file --output-format streaming-json`, stdin
+// ignored). Under load it would finish producing output but never terminate the
+// process (no controlling TTY to close on, an internal render/input wait, or a
+// tool it retried forever). `close` never fired → the per-message exit file was
+// never written → the task loop blocked until its 30 s sentinel / the 10 min
+// watchdog, and the turn surfaced as a StopFailure. The wedge root-cause was the
+// no-TTY non-exit.
 //
-// streaming-json lines are parsed; assistant text chunks are appended to
-// stdoutFile. exitFile is written when the process exits.
+// The tmux reimplementation gives grok a real PTY. One detached tmux session per
+// workspace runs `grok --no-alt-screen --always-approve` interactively; each
+// task PASTES its prompt into the live pane and detects turn-end by output
+// quiescence (recipe below). Context lives in the one long-running grok process
+// — the live session IS the resumed context, so we no longer `--resume` per task.
 //
-// Model: none is forced — grok uses the account's own default. Set an explicit,
-// VALID model via settings.grokModel to override (`grok models` lists them).
+// Contract preserved for taskLoop (unchanged): stream assistant/pane text to the
+// SAME per-message stdoutFile; write the exit code to the SAME per-message
+// exitFile exactly ONCE, only when the turn genuinely ends. Session id is still
+// minted-once and persisted under getSessionId/saveSessionId('grok-tui') so the
+// office/loop see continuity and a dead session can be relaunched with --resume.
+//
+// If tmux is not installed, grok-tui transparently FALLS BACK to the old
+// headless spawn so nothing regresses on boxes without tmux.
 // ---------------------------------------------------------------------------
 var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
     if (k2 === undefined) k2 = k;
@@ -56,14 +71,20 @@ var __importStar = (this && this.__importStar) || (function () {
 })();
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.isGrokTuiBusy = isGrokTuiBusy;
+exports.getGrokTuiLastActivity = getGrokTuiLastActivity;
+exports.forceIdleGrokTui = forceIdleGrokTui;
+exports.tmuxAvailable = tmuxAvailable;
 exports.sendGrokPrompt = sendGrokPrompt;
 exports.sendGrokCliPrompt = sendGrokCliPrompt;
 exports.sendGrokTuiPrompt = sendGrokTuiPrompt;
+exports.steerGrokTui = steerGrokTui;
 exports.closeGrokTuiSession = closeGrokTuiSession;
 exports.closeAllGrokTuiSessions = closeAllGrokTuiSessions;
 exports.detectGrokTuiRateLimit = detectGrokTuiRateLimit;
 const path = __importStar(require("path"));
 const fs = __importStar(require("fs"));
+const os = __importStar(require("os"));
+const crypto = __importStar(require("crypto"));
 const readline = __importStar(require("readline"));
 const child_process = __importStar(require("child_process"));
 const crypto_1 = require("crypto");
@@ -78,33 +99,47 @@ const GROK_BIN = process.env['GROK_BIN'] ?? (() => {
     // Prefer a grok on PATH; otherwise fall back to the default install location
     // (~/.grok/bin/grok) so the provider works without PATH changes.
     try {
-        const home = require('os').homedir();
-        const local = require('path').join(home, '.grok', 'bin', 'grok');
-        if (require('fs').existsSync(local)) {
+        const home = os.homedir();
+        const local = path.join(home, '.grok', 'bin', 'grok');
+        if (fs.existsSync(local)) {
             return local;
         }
     }
     catch { /* ignore */ }
     return 'grok';
 })();
+/** tmux binary — overridable for non-standard installs. */
+const TMUX_BIN = process.env['TMUX_BIN'] ?? 'tmux';
+// Env-tunable knobs. `envNum` accepts any finite >= 0; 0 disables that guard.
+const envNum = (key, def) => {
+    const n = Number(process.env[key]);
+    return Number.isFinite(n) && n >= 0 ? n : def;
+};
 // ---------------------------------------------------------------------------
 // Per-workspace state
 // ---------------------------------------------------------------------------
-/** Roots with an actively-running grok turn. */
+/** Roots with an actively-running grok turn (headless OR tmux). */
 const _busyRoots = new Set();
-/** Active child processes by root — used to kill them on extension deactivate. */
+/** Active HEADLESS child processes by root — grok-cli + tmux-unavailable fallback. */
 const _activeChildren = new Map();
+const _tmuxSessions = new Map();
+/** Epoch-ms of the last streamed pane growth per root (activity heartbeat). */
+const _lastActivityMs = new Map();
 /** True while a grok turn is running for the given workspace root. */
 function isGrokTuiBusy(root) {
     return _busyRoots.has(root);
 }
+/** Epoch-ms of the most recent streamed pane activity, or 0 if none. */
+function getGrokTuiLastActivity(root) {
+    return _lastActivityMs.get(root) ?? 0;
+}
+/** Force-clear the busy flag for a root whose turn appears hung. */
+function forceIdleGrokTui(root) {
+    _busyRoots.delete(root);
+}
 /**
  * Append a REAL grok activity event to `.autodev/hooks-events.jsonl` in the
- * native Claude-Code hook schema (pixel-office reads `hook_event_name`). Grok
- * has no hooks mechanism, but its `--output-format streaming-json` stream
- * carries tool_use / tool_result events — we translate those into PreToolUse /
- * PostToolUse so pixel-office shows real per-tool activity (no synthetic
- * SessionStart/End padding beyond the genuine turn boundaries).
+ * native Claude-Code hook schema (pixel-office reads `hook_event_name`).
  */
 function _emitGrokHook(root, hookEventName, extra = {}) {
     try {
@@ -112,29 +147,718 @@ function _emitGrokHook(root, hookEventName, extra = {}) {
         if (!fs.existsSync(dir)) {
             fs.mkdirSync(dir, { recursive: true });
         }
-        // Emit the canonical event_type too (grok writes hooks directly, bypassing
-        // normalizeEvent) so pixel-office can trust it without a provider map.
         const ev = { hook_event_name: hookEventName, event_type: (0, hookEventNormalizer_1.eventTypeFor)(hookEventName), provider: 'grok-tui', cwd: root, timestamp: new Date().toISOString(), ...extra };
         fs.appendFileSync(path.join(dir, 'hooks-events.jsonl'), JSON.stringify(ev) + '\n', 'utf8');
     }
     catch { /* non-critical */ }
 }
+// ---------------------------------------------------------------------------
+// tmux helpers — all synchronous, best-effort (swallow + report).
+// ---------------------------------------------------------------------------
+let _tmuxAvailable = null;
+/** True if tmux is installed and runnable. Cached after first probe. */
+function tmuxAvailable() {
+    if (_tmuxAvailable !== null) {
+        return _tmuxAvailable;
+    }
+    try {
+        child_process.execFileSync(TMUX_BIN, ['-V'], { stdio: 'ignore', timeout: 5_000 });
+        _tmuxAvailable = true;
+    }
+    catch {
+        _tmuxAvailable = false;
+    }
+    return _tmuxAvailable;
+}
+/** Run a tmux subcommand; returns stdout string, or null on any failure. */
+function tmux(args, input) {
+    try {
+        const out = child_process.execFileSync(TMUX_BIN, args, {
+            encoding: 'utf8', timeout: 15_000, input, stdio: ['pipe', 'pipe', 'ignore'],
+        });
+        return out ?? '';
+    }
+    catch {
+        return null;
+    }
+}
+/** Deterministic, tmux-safe session name from the absolute workspace path. */
+function sessionName(root) {
+    const h = crypto.createHash('sha1').update(root).digest('hex').slice(0, 10);
+    return `grok-${h}`;
+}
+/** True if the tmux session exists (a dead grok collapses the session). */
+function hasSession(name) {
+    try {
+        child_process.execFileSync(TMUX_BIN, ['has-session', '-t', name], { stdio: 'ignore', timeout: 5_000 });
+        return true;
+    }
+    catch {
+        return false;
+    }
+}
+/** Kill a tmux session (best-effort). */
+function killSession(name) {
+    tmux(['kill-session', '-t', name]);
+}
+/** Escape-stripped visible grid + scrollback of a pane (for detection). */
+function capturePane(name, scrollback = 200) {
+    const out = tmux(['capture-pane', '-p', '-t', name, '-S', `-${scrollback}`]);
+    return out === null ? '' : stripAnsi(out);
+}
+/** Shell-quote a single argument for a `send-keys "exec …"` command string. */
+function shq(s) {
+    return `'${s.replace(/'/g, `'\\''`)}'`;
+}
+/**
+ * Strip ANSI/OSC/DCS escapes and carriage returns from raw pane bytes.
+ * `pipe-pane` captures the full TUI byte stream (SGR colour, mouse tracking,
+ * bracketed-paste markers, OSC titles, `\r`), which must be removed before the
+ * text is written to stdoutFile or scanned for banners.
+ */
+function stripAnsi(s) {
+    return s
+        // CSI sequences: ESC [ ... final-byte
+        .replace(/\x1b\[[0-9;?]*[ -/]*[@-~]/g, '')
+        // OSC sequences: ESC ] ... (BEL | ST)
+        .replace(/\x1b\][^\x07\x1b]*(?:\x07|\x1b\\)/g, '')
+        // DCS/PM/APC: ESC (P|X|^|_) ... ST
+        .replace(/\x1b[PX^_][^\x1b]*\x1b\\/g, '')
+        // Any other single-char escape
+        .replace(/\x1b[@-Z\\-_]/g, '')
+        .replace(/\r/g, '');
+}
+// ---------------------------------------------------------------------------
+// Turn-end / auth detectors (calibrate against a live authed grok — see risks).
+// ---------------------------------------------------------------------------
+/**
+ * Markers that mean grok hit the OAuth/expired-token gate instead of a prompt.
+ * The loop must treat this as reauth, NOT a hung turn.
+ */
+const REAUTH_MARKERS = /waiting for approval|approve in your browser|sign(?:ing)? in to grok|finish signing in|device code|please (?:log ?in|sign ?in)/i;
+/**
+ * "grok finished this turn" affordance — CALIBRATED against grok 4.5 build 0.2.93:
+ * a completed turn prints "Worked for 4.1s." (a summary line) and the footer drops
+ * the cancel affordance. Used as a fast, positive turn-end confirmation.
+ */
+const DONE_MARKER = /worked for\s+[\d.]+\s*s\b/i;
+/**
+ * "grok is actively working" affordance. grok renders a live spinner + an
+ * "esc to interrupt"-style footer while a turn runs. If present we never
+ * declare idle. This is a SOFT guard: the primary idle signal is output
+ * quiescence + a byte-identical pane snapshot across polls (a running spinner
+ * animates, so the snapshot is never stable while grok works), which fires
+ * cleanly even if this pattern never matches on a given grok build.
+ */
+// CALIBRATED against real grok 4.5 (build 0.2.93): while a turn runs the footer
+// shows "Ctrl+c:cancel" and a "⠧ Waiting for response… 1.5s … [stop]" status line;
+// when idle the footer is only "Shift+Tab:mode │ Ctrl+x:shortcuts" (no Ctrl+c).
+// So the presence of Ctrl+c / "waiting for response" / "[stop]" / a braille spinner
+// = working. This is the primary running signal (backed by output-quiescence).
+const RUNNING_INDICATOR = /ctrl\+c\s*[: ]|waiting for response|\[stop\]|esc to (?:interrupt|cancel)|⠋|⠙|⠹|⠸|⠼|⠴|⠦|⠧|⠇|⠏/i;
+// ---------------------------------------------------------------------------
+// grok-tui: persistent interactive session driven through tmux.
+// ---------------------------------------------------------------------------
+/** Launch a fresh grok process in a new detached tmux session for this root. */
+function launchSession(root, sessionId, resume, model, log) {
+    const name = sessionName(root);
+    // A stale/dead session for this name must be cleared first.
+    if (hasSession(name)) {
+        killSession(name);
+    }
+    // Kill the one-time project-directory picker before it can block the turn.
+    ensureGrokPickerDisabled(log);
+    const dir = path.join(root, '.autodev', 'grok-tui');
+    try {
+        if (!fs.existsSync(dir)) {
+            fs.mkdirSync(dir, { recursive: true });
+        }
+    }
+    catch { /* ignore */ }
+    const rawLog = path.join(dir, 'pane.raw');
+    // Start each launch with a fresh raw log so readOffset math is simple.
+    try {
+        fs.writeFileSync(rawLog, '', 'utf8');
+    }
+    catch { /* ignore */ }
+    // Fixed geometry so wrapping/relayout never shifts detection rows; manual
+    // window-size so a human attaching can't reflow the pane mid-turn.
+    tmux(['new-session', '-d', '-s', name, '-x', '200', '-y', '50', '-c', root]);
+    tmux(['set-option', '-t', name, 'window-size', 'manual']);
+    // pipe-pane BEFORE launching grok — it only captures bytes emitted after it
+    // attaches (confirmed). Append mode.
+    tmux(['pipe-pane', '-o', '-t', name, `cat >> ${shq(rawLog)}`]);
+    // Build the interactive grok command. `exec` replaces the shell so a dead grok
+    // collapses the pane/session → cheap liveness signal via has-session.
+    const parts = [
+        'exec', shq(GROK_BIN),
+        '--no-alt-screen', '--always-approve',
+        '--cwd', shq(root),
+    ];
+    if (model) {
+        parts.push('-m', shq(model));
+    }
+    if (resume) {
+        // Restore the accumulated conversation on relaunch.
+        parts.push('--resume', shq(sessionId));
+    }
+    else {
+        // Name a brand-new conversation with our UUID so it is addressable later.
+        parts.push('-s', shq(sessionId));
+    }
+    tmux(['send-keys', '-t', name, parts.join(' '), 'Enter']);
+    const sess = { name, rawLog, sessionId, readOffset: 0 };
+    _tmuxSessions.set(root, sess);
+    log(`Grok TUI: launched tmux session ${name} (${resume ? 'resume' : 'new'} ${sessionId.slice(0, 8)}…, model=${model || 'account default'})`);
+    return sess;
+}
+/**
+ * Ensure a live tmux grok session exists for this root. Returns the session and
+ * whether it was just launched (caller applies a startup grace before pasting).
+ */
+function ensureSession(root, resolvedSessionId, model, log) {
+    const name = sessionName(root);
+    const cached = _tmuxSessions.get(root);
+    if (cached && cached.name === name && hasSession(name)) {
+        return { sess: cached, justLaunched: false }; // reuse — context is live in-process
+    }
+    // No live session. If we have a stored session id, relaunch with --resume so
+    // grok restores the accumulated history; otherwise mint + persist a new UUID.
+    let sid = resolvedSessionId || (0, sessionState_1.getSessionId)(root, 'grok-tui');
+    let resume = false;
+    if (sid) {
+        resume = true; // a stored id means grok already has a persisted session
+    }
+    else {
+        sid = (0, crypto_1.randomUUID)();
+        try {
+            (0, sessionState_1.saveSessionId)(root, 'grok-tui', sid);
+        }
+        catch { /* best effort */ }
+    }
+    const sess = launchSession(root, sid, resume, model, log);
+    return { sess, justLaunched: true };
+}
+/** Read new bytes of rawLog past `offset`; returns decoded text + new offset. */
+function readFrom(file, offset) {
+    try {
+        const size = fs.statSync(file).size;
+        if (size <= offset) {
+            return { text: '', offset };
+        }
+        const len = size - offset;
+        const buf = Buffer.alloc(len);
+        const fd = fs.openSync(file, 'r');
+        try {
+            fs.readSync(fd, buf, 0, len, offset);
+        }
+        finally {
+            fs.closeSync(fd);
+        }
+        return { text: buf.toString('utf8'), offset: size };
+    }
+    catch {
+        return { text: '', offset };
+    }
+}
+/**
+ * grok persists a CLEAN structured transcript per workspace session at
+ *   ~/.grok/sessions/<encodeURIComponent(cwd)>/<sessionId>/chat_history.jsonl
+ * one JSON object per line: {type:'system'|'user'|'reasoning'|'assistant'|
+ * 'tool_result', content, tool_calls?}. The assistant `content` is the real
+ * message text (markdown), with none of the TUI chrome — spinners, "Waiting for
+ * response…", token counters, "Ctrl+x:shortcuts", load-bar glyphs — that scraping
+ * the live tmux pane drags in. We read the agent's OUTPUT from here; the pane is
+ * used only for turn-end/reauth detection.
+ */
+function chatHistoryPath(root, sessionId) {
+    return path.join(os.homedir(), '.grok', 'sessions', encodeURIComponent(root), sessionId, 'chat_history.jsonl');
+}
 /**
- * Shared runner behind both grok providers. `persist:false` (grok-cli) spawns
- * a stateless process; `persist:true` (grok-tui) resumes the workspace session
- * (or creates + saves a new one on the first task).
+ * A NEW grok session shows a one-time "Run Grok Build in a project directory?"
+ * picker before the first turn's work, which blocks headless automation. grok
+ * exposes a persistent opt-out — `hints = { project_picker_disabled = true }` in
+ * ~/.grok/config.toml (what its "Don't ask me again" option writes). Ensure it is
+ * present before launching so the session goes straight to the prompt. Idempotent;
+ * the poll-loop picker backstop covers any grok build that ignores this.
  */
-function sendGrokPrompt(root, 
-/** Absolute path to the combined agent-profile + message file. */
-promptFilePath, stdoutFile, exitFile, log, opts = {}) {
+function ensureGrokPickerDisabled(log) {
+    try {
+        const cfg = path.join(os.homedir(), '.grok', 'config.toml');
+        let text = '';
+        try {
+            text = fs.readFileSync(cfg, 'utf8');
+        }
+        catch { /* no config yet */ }
+        if (/project_picker_disabled\s*=\s*true/.test(text)) {
+            return;
+        }
+        if (/^hints\s*=\s*\{/m.test(text)) {
+            // Inject the key into the existing inline `hints` table.
+            text = text.replace(/^hints\s*=\s*\{/m, 'hints = { project_picker_disabled = true,');
+        }
+        else {
+            text = 'hints = { project_picker_disabled = true }\n' + text;
+        }
+        fs.mkdirSync(path.dirname(cfg), { recursive: true });
+        fs.writeFileSync(cfg, text, 'utf8');
+        log('Grok TUI: disabled grok project-directory picker in config.toml');
+    }
+    catch { /* best effort — the poll-loop dismissal is the backstop */ }
+}
+/** Compact one-line marker for a grok tool call (for the live activity feed). */
+function formatToolCall(tc) {
+    const name = tc?.name || 'tool';
+    let arg = '';
+    try {
+        const a = JSON.parse(tc?.arguments || '{}');
+        const cand = a.target_file ?? a.file_path ?? a.path ?? a.command ?? a.query ?? a.message ?? a.to ?? '';
+        if (typeof cand === 'string') {
+            arg = cand;
+        }
+    }
+    catch { /* non-JSON args — just show the name */ }
+    arg = arg.replace(/\s+/g, ' ').trim().slice(0, 80);
+    return `» ${name}${arg ? ` ${arg}` : ''}`;
+}
+/**
+ * Read new chat_history.jsonl records past `offset` and render only the
+ * user-facing parts: assistant prose + compact tool-call markers. Skips
+ * reasoning/user/system/tool_result. Advances the offset only past COMPLETE
+ * lines so a half-written trailing record is re-read next poll.
+ */
+function drainChatHistory(file, offset) {
+    let size = 0;
+    try {
+        size = fs.statSync(file).size;
+    }
+    catch {
+        return { text: '', offset };
+    }
+    if (size <= offset) {
+        return { text: '', offset };
+    }
+    let raw = '';
+    try {
+        const len = size - offset;
+        const buf = Buffer.alloc(len);
+        const fd = fs.openSync(file, 'r');
+        try {
+            fs.readSync(fd, buf, 0, len, offset);
+        }
+        finally {
+            fs.closeSync(fd);
+        }
+        raw = buf.toString('utf8');
+    }
+    catch {
+        return { text: '', offset };
+    }
+    const lastNl = raw.lastIndexOf('\n');
+    if (lastNl < 0) {
+        return { text: '', offset };
+    } // no complete line yet
+    const complete = raw.slice(0, lastNl);
+    const newOffset = offset + Buffer.byteLength(complete, 'utf8') + 1; // +1 for the \n
+    let out = '';
+    for (const line of complete.split('\n')) {
+        const s = line.trim();
+        if (!s) {
+            continue;
+        }
+        let o;
+        try {
+            o = JSON.parse(s);
+        }
+        catch {
+            continue;
+        }
+        if ((o.type || o.role) !== 'assistant') {
+            continue;
+        }
+        const c = (o.content || '').trim();
+        if (c) {
+            out += c + '\n';
+        }
+        for (const tc of o.tool_calls || []) {
+            out += formatToolCall(tc) + '\n';
+        }
+    }
+    return { text: out, offset: newOffset };
+}
+/** Type a prompt into the live pane and submit it. CALIBRATED against real grok:
+ *  a tmux bracketed paste (paste-buffer) is misread by grok's TUI as a
+ *  worktree/directory PICKER, not chat input — only literal keystrokes land in
+ *  the input box. grok also has no reliable "newline without submit" key over
+ *  send-keys (Enter submits), so flatten the prompt to one logical line. */
+function pastePrompt(sess, promptFilePath) {
+    // A full agent prompt is 10KB+ (system prompt + task + context). Typing that
+    // via send-keys -l is UNRELIABLE (tmux drops/garbles keys at that size —
+    // observed live), and a bracketed paste is misread by grok's TUI as a directory
+    // picker. So DON'T type the prompt: hand grok a SHORT literal instruction to
+    // READ the prompt file, and its file tool ingests the whole thing losslessly.
+    const instr = `Read the file ${promptFilePath} in full right now — it is your current task with all its context and instructions. Do exactly what it says: do the real work, then mark the task done in TODO.md. Do not ask questions; do not skip the file.`;
+    tmux(['send-keys', '-t', sess.name, '-l', instr]);
+    tmux(['send-keys', '-t', sess.name, 'Enter']);
+}
+// ---------------------------------------------------------------------------
+// runGrokTmuxTurn — one turn against the persistent tmux session.
+// Fire-and-forget: streams pane text to stdoutFile, writes exitFile at turn end.
+// ---------------------------------------------------------------------------
+function runGrokTmuxTurn(root, promptFilePath, resolvedSessionId, stdoutFile, exitFile, log, model, showOutput) {
+    showOutput?.();
+    try {
+        fs.writeFileSync(stdoutFile, '', 'utf8');
+    }
+    catch { /* ignore */ }
+    try {
+        fs.writeFileSync(exitFile, '', 'utf8');
+    }
+    catch { /* ignore */ }
+    _busyRoots.add(root);
+    const POLL_MS = envNum('AUTODEV_GROK_TUI_POLL_MS', 700);
+    const DEBOUNCE_MS = envNum('AUTODEV_GROK_TUI_DEBOUNCE_MS', 5_000);
+    // Quiescence-only fallback (no "Worked for" marker seen) needs a much longer
+    // quiet window so grok's early "Starting session…"/"Reading file…" pauses aren't
+    // misread as turn-end (observed a false ~4s finish). The "Worked for" marker is
+    // the fast, positive done-signal for the normal case.
+    const QUIET_FALLBACK_MS = envNum('AUTODEV_GROK_TUI_QUIET_FALLBACK_MS', 45_000);
+    const STARTUP_MS = envNum('AUTODEV_GROK_TUI_STARTUP_MS', 8_000);
+    const NOOUTPUT_MS = envNum('AUTODEV_GROK_TUI_NOOUTPUT_MS', 25_000);
+    const MAX_RUN_MS = envNum('AUTODEV_GROK_MAX_RUN_MS', 10 * 60_000);
+    const STABLE_POLLS = Math.max(2, envNum('AUTODEV_GROK_TUI_STABLE_POLLS', 4));
+    const sleep = (ms) => new Promise(r => setTimeout(r, ms));
+    // Line-buffer for clean console logging (grok streams mid-word).
+    let lineBuf = '';
+    const emitLines = (text) => {
+        lineBuf += text;
+        let nl;
+        while ((nl = lineBuf.indexOf('\n')) >= 0) {
+            const line = lineBuf.slice(0, nl).replace(/\s+$/, '');
+            lineBuf = lineBuf.slice(nl + 1);
+            if (line.trim()) {
+                log(`  ${line}`);
+            }
+        }
+    };
+    const flushLine = () => {
+        const line = lineBuf.replace(/\s+$/, '');
+        lineBuf = '';
+        if (line.trim()) {
+            log(`  ${line}`);
+        }
+    };
+    // Coalesced office activity feed (grok exposes no per-tool events here).
+    let activityBuf = '';
+    let activityTimer = null;
+    const flushActivity = () => {
+        if (activityTimer) {
+            clearTimeout(activityTimer);
+            activityTimer = null;
+        }
+        const t = activityBuf.trim();
+        activityBuf = '';
+        if (!t) {
+            return;
+        }
+        const preview = t.length > 280 ? t.slice(0, 277) + '…' : t;
+        _emitGrokHook(root, 'Notification', { message: preview, title: 'grok', tool_name: 'grok' });
+    };
+    const scheduleActivity = () => {
+        if (activityBuf.length >= 400) {
+            flushActivity();
+            return;
+        }
+        if (!activityTimer) {
+            activityTimer = setTimeout(flushActivity, 1200);
+        }
+    };
+    const finish = (code, reason) => {
+        flushLine();
+        flushActivity();
+        _busyRoots.delete(root);
+        log(`Grok TUI: turn ${code === 0 ? 'complete' : `ended (code=${code}, ${reason})`}`);
+        _emitGrokHook(root, code === 0 ? 'Stop' : 'StopFailure', { exit_code: code });
+        _emitGrokHook(root, 'SessionEnd', { reason: code === 0 ? 'completed' : reason });
+        try {
+            fs.writeFileSync(exitFile, `${code}\n`, 'utf8');
+        }
+        catch { /* ignore */ }
+    };
+    void (async () => {
+        let sess;
+        let justLaunched;
+        try {
+            const r = ensureSession(root, resolvedSessionId, model, log);
+            sess = r.sess;
+            justLaunched = r.justLaunched;
+        }
+        catch (err) {
+            const msg = err?.message ?? String(err);
+            log(`Grok TUI: session launch failed: ${msg}`);
+            try {
+                fs.appendFileSync(stdoutFile, `\n[Grok TUI launch error: ${msg}]\n`, 'utf8');
+            }
+            catch { /* ignore */ }
+            finish(1, 'launch-failed');
+            return;
+        }
+        // Stream only THIS turn's output: start reading at the current rawLog size.
+        let readOffset = (() => { try {
+            return fs.statSync(sess.rawLog).size;
+        }
+        catch {
+            return 0;
+        } })();
+        if (justLaunched) {
+            _emitGrokHook(root, 'SessionStart', { source: 'startup' });
+            // Startup grace — grok's splash takes a few seconds before it accepts input.
+            // Scan for the reauth gate while we wait; bail early if the token is dead.
+            const deadline = Date.now() + STARTUP_MS;
+            while (Date.now() < deadline) {
+                await sleep(500);
+                if (!hasSession(sess.name)) {
+                    log('Grok TUI: session died during startup');
+                    try {
+                        fs.appendFileSync(stdoutFile, `\n[Grok TUI: session exited during startup]\n`, 'utf8');
+                    }
+                    catch { /* ignore */ }
+                    finish(1, 'startup-exit');
+                    return;
+                }
+                const snap = capturePane(sess.name);
+                if (REAUTH_MARKERS.test(snap)) {
+                    log('Grok TUI: reauth required (OAuth gate at startup)');
+                    try {
+                        fs.appendFileSync(stdoutFile, `\n[Grok TUI: reauthentication required — please login]\n`, 'utf8');
+                    }
+                    catch { /* ignore */ }
+                    finish(1, 'reauth_required');
+                    return;
+                }
+            }
+            // CALIBRATED: a NEW grok session opens on a picker sitting over the input
+            // box — either the welcome menu (New worktree / Resume session / Changelog /
+            // Quit) or the project chooser ("Run Grok Build in a project directory?"
+            // with radio options, the current dir highlighted). A RESUMED session skips
+            // it. Pressing Enter confirms the highlighted default (current dir / first
+            // item) and proceeds to the ready "❯" prompt. A single fixed Enter races the
+            // picker's render (grok's splash can outlast the grace), so poll and press
+            // Enter each time a picker is still showing, until the prompt is clear.
+            const PICKER_MARKERS = /project directory|New worktree|Resume session|Changelog|\(current\)|Run Grok Build|[◯○◉●]|Select|↑\/↓/i;
+            for (let i = 0; i < 8; i++) {
+                if (!hasSession(sess.name)) {
+                    log('Grok TUI: session died during startup');
+                    try {
+                        fs.appendFileSync(stdoutFile, `\n[Grok TUI: session exited during startup]\n`, 'utf8');
+                    }
+                    catch { /* ignore */ }
+                    finish(1, 'startup-exit');
+                    return;
+                }
+                const snap = capturePane(sess.name);
+                if (REAUTH_MARKERS.test(snap)) {
+                    log('Grok TUI: reauth required (OAuth gate at startup)');
+                    try {
+                        fs.appendFileSync(stdoutFile, `\n[Grok TUI: reauthentication required — please login]\n`, 'utf8');
+                    }
+                    catch { /* ignore */ }
+                    finish(1, 'reauth_required');
+                    return;
+                }
+                if (!PICKER_MARKERS.test(snap)) {
+                    break;
+                } // at the ready prompt
+                tmux(['send-keys', '-t', sess.name, 'Enter']);
+                await sleep(1500);
+            }
+            // re-read the offset after startup so the splash/menu text isn't streamed
+            // as "turn output".
+            readOffset = (() => { try {
+                return fs.statSync(sess.rawLog).size;
+            }
+            catch {
+                return readOffset;
+            } })();
+        }
+        else {
+            _emitGrokHook(root, 'SessionStart', { source: 'resume' });
+        }
+        // Clean-output source: grok's structured transcript. Seed the offset at the
+        // current size so we emit only THIS turn's assistant records, not the
+        // resumed history. (File may not exist yet on a fresh session → offset 0.)
+        const chatPath = chatHistoryPath(root, sess.sessionId);
+        let chatOffset = (() => { try {
+            return fs.statSync(chatPath).size;
+        }
+        catch {
+            return 0;
+        } })();
+        let chatEmitted = false;
+        // Submit the prompt.
+        log(`Grok TUI: sending turn (${(() => { try {
+            return fs.statSync(promptFilePath).size;
+        }
+        catch {
+            return 0;
+        } })()} bytes)`);
+        pastePrompt(sess, promptFilePath);
+        // -----------------------------------------------------------------------
+        // Poll: stream new pane bytes, detect turn-end by output quiescence + a
+        // byte-stable pane snapshot with the running-indicator absent.
+        // -----------------------------------------------------------------------
+        const turnStart = Date.now();
+        let lastGrowthMs = Date.now();
+        let sawOutput = false;
+        let lastSnap = '';
+        let stable = 0;
+        let sawDoneMarker = false;
+        for (;;) {
+            await sleep(POLL_MS);
+            // Session died mid-turn (grok crashed / was killed).
+            if (!hasSession(sess.name)) {
+                try {
+                    fs.appendFileSync(stdoutFile, `\n[Grok TUI: session exited]\n`, 'utf8');
+                }
+                catch { /* ignore */ }
+                _tmuxSessions.delete(root);
+                finish(1, 'session-exit');
+                return;
+            }
+            // 1a) Read the pane ONLY for activity/quiescence timing — its bytes are
+            //     TUI chrome (spinners, "Waiting for response…", token counters) and
+            //     must NOT be emitted as the agent's output.
+            const { text: paneText, offset: paneOff } = readFrom(sess.rawLog, readOffset);
+            if (paneText) {
+                readOffset = paneOff;
+                sess.readOffset = paneOff;
+                if (stripAnsi(paneText).trim()) {
+                    _lastActivityMs.set(root, Date.now());
+                    lastGrowthMs = Date.now();
+                    sawOutput = true;
+                }
+            }
+            // 1b) Emit CLEAN output from grok's structured transcript (assistant prose
+            //     + compact tool markers), never the pane scrape.
+            const ch = drainChatHistory(chatPath, chatOffset);
+            if (ch.text) {
+                chatOffset = ch.offset;
+                chatEmitted = true;
+                try {
+                    fs.appendFileSync(stdoutFile, ch.text, 'utf8');
+                }
+                catch { /* ignore */ }
+                emitLines(ch.text);
+                activityBuf += ch.text;
+                scheduleActivity();
+                _lastActivityMs.set(root, Date.now());
+            }
+            // 2) Snapshot for detection (also catches a mid-turn reauth prompt).
+            const snap = capturePane(sess.name);
+            if (REAUTH_MARKERS.test(snap)) {
+                try {
+                    fs.appendFileSync(stdoutFile, `\n[Grok TUI: reauthentication required — please login]\n`, 'utf8');
+                }
+                catch { /* ignore */ }
+                finish(1, 'reauth_required');
+                return;
+            }
+            // Backstop: the project-directory picker appears AFTER the first prompt is
+            // submitted (grok's one-time confirm), so it lands mid-turn. The config
+            // opt-out normally prevents it; if a grok build ignores that, select the
+            // current dir (option "1", always present) to unblock the turn.
+            if (/Run Grok Build in a project directory|Don't ask me again|\(current\)/i.test(snap)) {
+                tmux(['send-keys', '-t', sess.name, '1']);
+                await sleep(800);
+                lastGrowthMs = Date.now();
+                continue;
+            }
+            // 3) Idle test. grok prints "Worked for N.Ns." when a turn genuinely ends —
+            //    that's the fast, positive done-signal. Absent it (an error turn, or an
+            //    unusual grok build), fall back to a MUCH longer quiescence so early
+            //    "Starting session…"/"Reading file…" pauses aren't misread as turn-end.
+            if (DONE_MARKER.test(snap)) {
+                sawDoneMarker = true;
+            }
+            if (snap === lastSnap) {
+                stable++;
+            }
+            else {
+                stable = 0;
+                lastSnap = snap;
+            }
+            const running = RUNNING_INDICATOR.test(snap);
+            const quietMs = Date.now() - lastGrowthMs;
+            const readyToJudge = sawOutput || (Date.now() - turnStart >= NOOUTPUT_MS);
+            const doneFast = sawDoneMarker && quietMs >= DEBOUNCE_MS;
+            const doneSlow = readyToJudge && quietMs >= QUIET_FALLBACK_MS && stable >= STABLE_POLLS;
+            if (!running && (doneFast || doneSlow)) {
+                // Final drain of the transcript to catch records written between the last
+                // poll and turn-end.
+                const tail = drainChatHistory(chatPath, chatOffset);
+                if (tail.text) {
+                    chatOffset = tail.offset;
+                    chatEmitted = true;
+                    try {
+                        fs.appendFileSync(stdoutFile, tail.text, 'utf8');
+                    }
+                    catch { /* ignore */ }
+                    emitLines(tail.text);
+                }
+                // Safety net: if the transcript path never resolved (grok layout change /
+                // encoding drift) yet the pane clearly produced output, fall back to a
+                // cleaned pane tail so the turn is never silently empty.
+                if (!chatEmitted && sawOutput) {
+                    const pane = stripAnsi(capturePane(sess.name, 400)).replace(/[ \t]+\n/g, '\n').trim();
+                    if (pane) {
+                        try {
+                            fs.appendFileSync(stdoutFile, pane + '\n', 'utf8');
+                        }
+                        catch { /* ignore */ }
+                    }
+                    log('Grok TUI: transcript yielded no output — fell back to pane capture (check chat_history path)');
+                }
+                finish(0, 'idle');
+                return;
+            }
+            // 4) Watchdog — a turn that overruns the hard budget. Interrupt with Esc
+            //    (grok's cancel) to preserve the session/context; if it won't settle,
+            //    kill the session so the NEXT turn relaunches with --resume.
+            if (MAX_RUN_MS > 0 && Date.now() - turnStart >= MAX_RUN_MS) {
+                log(`Grok TUI watchdog: exceeded ${Math.round(MAX_RUN_MS / 60_000)}min budget — interrupting`);
+                try {
+                    fs.appendFileSync(stdoutFile, `\n[Grok TUI watchdog: turn exceeded time budget — interrupted]\n`, 'utf8');
+                }
+                catch { /* ignore */ }
+                tmux(['send-keys', '-t', sess.name, 'Escape']);
+                // Give grok a moment to settle after the interrupt.
+                let settled = false;
+                for (let i = 0; i < 8; i++) {
+                    await sleep(500);
+                    if (!hasSession(sess.name)) {
+                        break;
+                    }
+                    const s2 = capturePane(sess.name);
+                    if (!RUNNING_INDICATOR.test(s2)) {
+                        settled = true;
+                        break;
+                    }
+                }
+                if (!settled && hasSession(sess.name)) {
+                    // Still wedged after Esc → kill so ensureSession relaunches next turn.
+                    killSession(sess.name);
+                    _tmuxSessions.delete(root);
+                }
+                finish(124, 'watchdog');
+                return;
+            }
+        }
+    })();
+}
+function sendGrokPrompt(root, promptFilePath, stdoutFile, exitFile, log, opts = {}) {
     opts.showOutput?.();
-    // Only force a model when one is explicitly configured. With no model, grok
-    // uses the account's own default — more robust than hardcoding a model id
-    // that may not exist for every account/plan.
     const modelArgs = opts.model ? ['-m', opts.model] : [];
     const modelLabel = opts.model || 'account default';
-    // Session flags — grok-tui only. Resume the stored session, or mint a new id
-    // and persist it immediately so the next task resumes the same conversation.
     const sessionArgs = [];
     if (opts.persist) {
         const providerId = opts.providerId ?? 'grok-tui';
@@ -163,9 +887,6 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
         fs.writeFileSync(exitFile, '', 'utf8');
     }
     catch { /* ignore */ }
-    // Defensive: if a previous run for this root is somehow still tracked, kill it
-    // first so `_activeChildren.set` below can't orphan it (callers gate on
-    // isGrokTuiBusy, but don't rely on that alone).
     const prior = _activeChildren.get(root);
     if (prior) {
         try {
@@ -175,17 +896,12 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
         _activeChildren.delete(root);
     }
     _busyRoots.add(root);
-    /**
-     * Fail the turn the way the (previously unreachable) catch below intended.
-     * Shared so the sync and async paths can't drift into disagreeing.
-     */
     const failSpawn = (msg) => {
         log(`Grok spawn error: ${msg}`);
         try {
             fs.appendFileSync(stdoutFile, `\n[Grok spawn error: ${msg}]\n`, 'utf8');
         }
         catch { /* ignore */ }
-        // The exit file is what the loop waits on — without it the turn never ends.
         try {
             fs.writeFileSync(exitFile, '1\n', 'utf8');
         }
@@ -202,28 +918,12 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
             ...sessionArgs,
             '--prompt-file', promptFilePath,
             '--output-format', 'streaming-json',
-        ], {
-            cwd: root,
-            stdio: ['ignore', 'pipe', 'pipe'],
-            env: { ...process.env },
-        });
+        ], { cwd: root, stdio: ['ignore', 'pipe', 'pipe'], env: { ...process.env } });
     }
     catch (spawnErr) {
-        // Kept for the genuinely synchronous failures (bad options/EACCES on some
-        // platforms). A MISSING BINARY does NOT land here — see the 'error' handler.
         failSpawn(spawnErr?.message ?? String(spawnErr));
         return;
     }
-    // The one that actually fires when grok isn't installed.
-    //
-    // `spawn()` does NOT throw on ENOENT — it emits 'error' asynchronously — so the
-    // catch above was dead code for the most common failure by far. With no handler
-    // the error escaped to start.ts's uncaughtException backstop, which keeps the
-    // process alive: 'close' never fired, the exit file stayed empty, _busyRoots was
-    // never cleared, and SessionStart had ALREADY been emitted — so the office showed
-    // the agent WORKING, forever, on a grok that was never running. Observed live:
-    // an agent "online" and busy for hours whose only trace was `spawn grok ENOENT`
-    // buried in agent.log. (opencodeCliProvider always did this correctly.)
     child.on('error', (err) => {
         const missing = err.code === 'ENOENT';
         failSpawn(missing
@@ -231,23 +931,7 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
             : err.message);
     });
     _activeChildren.set(root, child);
-    // Emit SessionStart only AFTER the child is confirmed running. Emitting it
-    // before spawn resolves is what let a failed turn masquerade as an active one.
     child.once('spawn', () => { _emitGrokHook(root, 'SessionStart', { source: 'startup' }); });
-    // -------------------------------------------------------------------------
-    // Watchdog — grok can wedge internally (e.g. retrying a failing tool call
-    // forever) WITHOUT ever exiting, which leaves the process alive, the exit
-    // file unwritten, and the task loop blocked indefinitely on this one task
-    // while new work piles up. Kill a run that overruns a hard time budget or
-    // floods tool-output errors, so `close` fires, the exit file is written, and
-    // the loop fails this task cleanly and moves on.
-    // -------------------------------------------------------------------------
-    // Env overrides: accept any finite >= 0; 0 explicitly DISABLES that guard
-    // (the old `Number(x) || DEFAULT` silently reverted 0 back to the default).
-    const envNum = (key, def) => {
-        const n = Number(process.env[key]);
-        return Number.isFinite(n) && n >= 0 ? n : def;
-    };
     const GROK_MAX_RUN_MS = envNum('AUTODEV_GROK_MAX_RUN_MS', 10 * 60_000);
     const GROK_MAX_TOOL_ERRORS = envNum('AUTODEV_GROK_MAX_TOOL_ERRORS', 25);
     let toolErrorCount = 0;
@@ -255,7 +939,7 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
     const killGrok = (why) => {
         if (killed) {
             return;
-        } // never double-kill / double-log
+        }
         killed = true;
         log(`Grok watchdog: ${why} — killing run`);
         try {
@@ -270,21 +954,8 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
     const watchdog = GROK_MAX_RUN_MS > 0
         ? setTimeout(() => killGrok(`exceeded ${Math.round(GROK_MAX_RUN_MS / 60_000)}min time budget`), GROK_MAX_RUN_MS)
         : null;
-    // -------------------------------------------------------------------------
-    // Stream stdout — each line is a streaming-json object.
-    // -------------------------------------------------------------------------
     const rl = readline.createInterface({ input: child.stdout });
-    // Grok's streaming-json exposes `thought`, `text`, `end` — but NOT tool
-    // events (tool use is hidden under --always-approve). So per-tool activity
-    // isn't available; we surface the real turn boundaries instead: SessionStart
-    // at spawn and Stop/SessionEnd on the `end` event (a genuine grok signal, not
-    // synthetic padding). `_endSeen` avoids a duplicate SessionEnd on close.
     let _endSeen = false;
-    // Grok exposes NO tool events in streaming-json (tool use is hidden under
-    // --always-approve), so pixel-office's activity feed would otherwise be empty
-    // apart from the SessionStart/End boundaries. Surface the assistant's streamed
-    // TEXT as periodic, coalesced `Notification` hook events so the office shows
-    // what grok is actually producing (not raw per-chunk spam).
     let _activityBuf = '';
     let _activityTimer = null;
     const flushActivity = () => {
@@ -309,10 +980,6 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
             _activityTimer = setTimeout(flushActivity, 1200);
         }
     };
-    // Console output: grok streams assistant text token-by-token (often mid-word,
-    // e.g. "aut"/"ode"/"v"). Logging each delta puts one timestamped line per token
-    // and shreds the output. Buffer deltas and only log COMPLETE lines (on \n),
-    // flushing whatever remains when the turn ends.
     let _lineBuf = '';
     const emitOutputLines = (text) => {
         _lineBuf += text;
@@ -341,7 +1008,6 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
             msg = JSON.parse(line);
         }
         catch {
-            // Plain text fallback (e.g. progress lines before JSON kicks in).
             try {
                 fs.appendFileSync(stdoutFile, line + '\n', 'utf8');
             }
@@ -350,16 +1016,13 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
         }
         const type = msg.type ?? '';
         if (type === 'assistant' || type === 'text') {
-            // Assistant response text.
             const text = msg.content ?? msg.data ?? msg.text ?? msg.message ?? '';
             if (text) {
                 try {
                     fs.appendFileSync(stdoutFile, text, 'utf8');
                 }
                 catch { /* ignore */ }
-                // Log full lines only (buffer partial tokens until a newline arrives).
                 emitOutputLines(text);
-                // Feed the office activity stream (coalesced).
                 _activityBuf += text;
                 scheduleActivity();
             }
@@ -373,17 +1036,13 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
             catch { /* ignore */ }
         }
         else if (type === 'end') {
-            // Genuine turn-end from grok's stream → real session boundary.
             _endSeen = true;
-            flushOutputLine(); // print any half-line still buffered
-            flushActivity(); // emit any buffered assistant text first
+            flushOutputLine();
+            flushActivity();
             _emitGrokHook(root, 'Stop', {});
             _emitGrokHook(root, 'SessionEnd', { reason: 'completed' });
         }
-        // `thought` and other event types stay out of the output file.
     });
-    // Read stderr LINE by line (not raw chunks) so a `tool_output_error` marker
-    // split across two data chunks is still counted exactly once.
     const errRl = child.stderr ? readline.createInterface({ input: child.stderr }) : null;
     errRl?.on('line', (line) => {
         const text = line.trim();
@@ -391,8 +1050,6 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
             return;
         }
         log(`Grok stderr: ${text}`);
-        // A flood of tool-output errors means grok is stuck retrying a tool it
-        // can't complete — kill it now rather than waiting out the time budget.
         if (GROK_MAX_TOOL_ERRORS > 0 && text.includes('tool_output_error')) {
             toolErrorCount++;
             if (toolErrorCount >= GROK_MAX_TOOL_ERRORS) {
@@ -404,16 +1061,14 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
         if (watchdog) {
             clearTimeout(watchdog);
         }
-        flushOutputLine(); // print any half-line still buffered
-        flushActivity(); // flush any trailing assistant text
+        flushOutputLine();
+        flushActivity();
         errRl?.close();
         rl.close();
         _activeChildren.delete(root);
         _busyRoots.delete(root);
         const exitCode = code ?? 1;
         log(`Grok: exited (code=${exitCode})`);
-        // Fallback SessionEnd only if grok didn't already emit an `end` event
-        // (abnormal exit / killed) — avoids a duplicate from the normal path.
         if (!_endSeen) {
             _emitGrokHook(root, exitCode === 0 ? 'Stop' : 'StopFailure', { exit_code: exitCode });
             _emitGrokHook(root, 'SessionEnd', { reason: exitCode === 0 ? 'completed' : 'error' });
@@ -424,28 +1079,78 @@ promptFilePath, stdoutFile, exitFile, log, opts = {}) {
         catch { /* ignore */ }
     });
 }
-/** grok-cli: stateless — a fresh grok process every task (no session flags). */
+/** grok-cli: stateless — a fresh HEADLESS grok process every task. */
 function sendGrokCliPrompt(root, promptFilePath, stdoutFile, exitFile, log, model, showOutput) {
     sendGrokPrompt(root, promptFilePath, stdoutFile, exitFile, log, {
         model, showOutput, persist: false, providerId: 'grok-cli',
     });
 }
-/** grok-tui: persistent — resume the workspace session so context accumulates. */
-function sendGrokTuiPrompt(root, promptFilePath, 
-/** Session id from resolveSession (undefined on the first task). */
-resolvedSessionId, stdoutFile, exitFile, log, model, showOutput) {
+/**
+ * grok-tui: persistent — run the turn in the per-workspace tmux session so
+ * grok's in-process context accumulates across tasks. Falls back to the old
+ * headless spawn (with --resume) when tmux is unavailable so nothing regresses.
+ */
+function sendGrokTuiPrompt(root, promptFilePath, resolvedSessionId, stdoutFile, exitFile, log, model, showOutput) {
+    if (tmuxAvailable()) {
+        runGrokTmuxTurn(root, promptFilePath, resolvedSessionId, stdoutFile, exitFile, log, model, showOutput);
+        return;
+    }
+    log('Grok TUI: tmux not available — falling back to headless spawn');
     sendGrokPrompt(root, promptFilePath, stdoutFile, exitFile, log, {
         model, showOutput, persist: true, sessionId: resolvedSessionId, providerId: 'grok-tui',
     });
 }
 // ---------------------------------------------------------------------------
-// closeGrokTuiSession
-// Called by taskLoop reset interval. grok-tui persists its session id in
-// session-state.json BY DESIGN (so context survives), so we only ensure no
-// child is left running — the session id is intentionally preserved.
+// steerGrokTui — inject a message into the RUNNING grok turn via the live pane.
+//
+// Because the tmux session is a real TTY, keys sent to a mid-turn pane are
+// delivered to grok's input and folded into the current turn (confirmed with a
+// live REPL stand-in). Mirrors steerClaudeTui: only injects when a turn is
+// actually running; returns false when there is no live pane so the caller
+// keeps the durable TODO fallback (at-least-once delivery).
+// ---------------------------------------------------------------------------
+async function steerGrokTui(root, text, log) {
+    const sess = _tmuxSessions.get(root);
+    if (!sess) {
+        return false;
+    } // no live tmux session (headless / not started)
+    if (!_busyRoots.has(root)) {
+        return false;
+    } // only steer a turn in flight
+    if (!hasSession(sess.name)) {
+        _tmuxSessions.delete(root);
+        return false;
+    }
+    try {
+        // Literal keystrokes (send-keys -l), NOT a bracketed paste — grok's TUI
+        // misreads a paste as a directory picker. Flatten to one line (no reliable
+        // newline-without-submit key), then Enter to inject into the running turn.
+        const flat = text.replace(/\r?\n+/g, ' ').replace(/[ \t]+/g, ' ').trim();
+        for (let i = 0; i < flat.length; i += 3000) {
+            if (tmux(['send-keys', '-t', sess.name, '-l', flat.slice(i, i + 3000)]) === null) {
+                return false;
+            }
+        }
+        if (tmux(['send-keys', '-t', sess.name, 'Enter']) === null) {
+            return false;
+        }
+        log(`Grok TUI: steered live pane (${text.length} chars) — injected into current turn`);
+        return true;
+    }
+    catch (err) {
+        log(`Grok TUI: steer failed: ${err?.message ?? String(err)}`);
+        return false;
+    }
+}
+// ---------------------------------------------------------------------------
+// Teardown
 // ---------------------------------------------------------------------------
+/**
+ * Called by taskLoop reset interval / provider close. Kills any live grok for
+ * this root — the tmux session AND any headless child — while PRESERVING the
+ * persisted session id so a later task relaunches with --resume (context intact).
+ */
 function closeGrokTuiSession(root, _log) {
-    // Kill any active child for this root; keep the persisted session id.
     const child = _activeChildren.get(root);
     if (child) {
         try {
@@ -454,9 +1159,21 @@ function closeGrokTuiSession(root, _log) {
         catch { /* ignore */ }
         _activeChildren.delete(root);
     }
+    const sess = _tmuxSessions.get(root);
+    if (sess) {
+        killSession(sess.name);
+        _tmuxSessions.delete(root);
+    }
+    else {
+        // No cached handle but a session may still exist under the deterministic name.
+        const name = sessionName(root);
+        if (hasSession(name)) {
+            killSession(name);
+        }
+    }
     _busyRoots.delete(root);
 }
-/** Kill all active grok processes — called on extension deactivate. */
+/** Kill all live grok sessions — called on SDK/extension shutdown. */
 function closeAllGrokTuiSessions() {
     for (const child of _activeChildren.values()) {
         try {
@@ -465,6 +1182,10 @@ function closeAllGrokTuiSessions() {
         catch { /* ignore */ }
     }
     _activeChildren.clear();
+    for (const sess of _tmuxSessions.values()) {
+        killSession(sess.name);
+    }
+    _tmuxSessions.clear();
     _busyRoots.clear();
 }
 // ---------------------------------------------------------------------------