Streaming overlay: chat GUI for agent responses

- overlay.zig: frameless translucent panel (fixed bg + rounded corners,
  only bubbles scroll), drag handle, hover-to-full-opacity, top-right spawn
- chat layout: per-message QLabel bubbles (user purple / agent green /
  tool steps grey mono), transcript retained across turns, End-conversation
  resets
- agent reply streams into its own growing bubble (re-render per chunk)
- basic markdown in all messages: **bold**, *italic*, `code`, HTML-escaped
- tool steps truncated to 90 chars, click to expand/collapse
- polite autoscroll: pins to bottom unless user scrolled up
- letta.zig: streamInfer() — --output-format stream-json, per-event callback
  (tool calls, returns, reply chunks), full reply still returned for TTS
- whisper: --language auto (no more hardcoded English)
- fix: stderr_log leak crashing Quit via debug allocator

👾 Generated with [Letta Code](https://letta.com)

Co-Authored-By: Letta Code <noreply@letta.com>

src/main.zig

@@ -19,6 +19,7 @@ const QBrush = qt6.QBrush;
 const QRectF = qt6.QRectF;
 const threading = qt6.threading;
 const letta = @import("letta.zig");
+const overlay_mod = @import("overlay.zig");
 
 // ------------------------------------------------------------- constants ---
 
@@ -82,6 +83,49 @@ var current_pixmap: ?QPixmap = null;
 var main_thread_id: std.Thread.Id = 0;
 var pending_state: State = .idle;
 
+// --------------------------------------------------------------- overlay ---
+
+var overlay: ?overlay_mod.Overlay = null;
+
+const OverlayKind = enum { begin, user, step, chunk };
+const OverlayMsg = struct { kind: OverlayKind, text: ?[]u8 };
+
+/// Worker-thread side: queue an overlay update onto the Qt main loop.
+fn pushOverlay(kind: OverlayKind, text: ?[]u8) void {
+    const msg = allocator.create(OverlayMsg) catch {
+        if (text) |t| allocator.free(t);
+        return;
+    };
+    msg.* = .{ .kind = kind, .text = text };
+    threading.async(msg, onOverlayMsg);
+}
+
+fn onOverlayMsg(ctx: ?*anyopaque) callconv(.c) void {
+    const msg: *OverlayMsg = @ptrCast(@alignCast(ctx orelse return));
+    defer allocator.destroy(msg);
+    if (overlay) |*ov| switch (msg.kind) {
+        .begin => ov.begin(),
+        .user => ov.user(msg.text orelse ""),
+        .step => ov.step(msg.text orelse ""),
+        .chunk => ov.chunk(msg.text orelse ""),
+    };
+    if (msg.text) |t| allocator.free(t);
+}
+
+
+fn onLettaEvent(_: ?*anyopaque, ev: letta.Event) void {
+    switch (ev) {
+        .step => |s| {
+            const dup = allocator.dupe(u8, s) catch return;
+            pushOverlay(.step, dup);
+        },
+        .chunk => |c| {
+            const dup = allocator.dupe(u8, c) catch return;
+            pushOverlay(.chunk, dup);
+        },
+    }
+}
+
 /// Must run on the Qt main thread (touches QSystemTrayIcon / QPixmap).
 fn setStateMainThread(next: State) void {
     state = next;
@@ -188,6 +232,7 @@ fn startRecording() void {
     // pw-record finalizes the WAV header on SIGTERM — kill() sends it.
     // stderr goes to a log so failures aren't silent.
     const stderr_log = std.fmt.allocPrint(allocator, "{s}/pw-record.log", .{RECORDINGS_DIR}) catch null;
+    defer if (stderr_log) |sl| allocator.free(sl);
     const stderr_file: ?Io.File = if (stderr_log) |lp|
         Io.Dir.cwd().createFile(io, lp[1..], .{ .truncate = true }) catch null
     else
@@ -265,7 +310,10 @@ fn processUtterance() void {
 
     // --- agent ---
     setState(.thinking);
-    const response = letta.infer(io, allocator, LETTA_AGENT_ID, text) catch |err| {
+
+    pushOverlay(.begin, null);
+    if (allocator.dupe(u8, text) catch null) |user_dup| pushOverlay(.user, user_dup);
+    const response = letta.streamInfer(io, allocator, LETTA_AGENT_ID, text, null, onLettaEvent) catch |err| {
         std.log.err("letta failed: {s}", .{@errorName(err)});
         setState(.idle);
         return;
@@ -323,6 +371,10 @@ fn onEndConversation(_: QAction) callconv(.c) void {
     conversation_active = false;
     if (last_user_text) |t| allocator.free(t);
     last_user_text = null;
+    if (overlay) |*ov| {
+        ov.reset();
+        ov.hide();
+    }
     setState(.idle);
 }
 
@@ -401,6 +453,10 @@ pub fn main(init: std.process.Init) !void {
     tray_icon.show();
     main_thread_id = std.Thread.getCurrentId();
 
+    var ov = overlay_mod.Overlay.create(allocator);
+    overlay = ov;
+    _ = &ov;
+
     // Own the whisper-server lifecycle: spawn AFTER the tray exists (setState
     // touches tray_icon). Create the log dir first.
     Io.Dir.createDirAbsolute(init.io, RECORDINGS_DIR, .default_dir) catch {};