the rest of the refactor

cht/session.py

@@ -0,0 +1,35 @@
+"""Session data loading — reads frame/transcript indexes, returns plain data."""
+
+import json
+import logging
+from pathlib import Path
+
+log = logging.getLogger(__name__)
+
+
+def load_frame_index(frames_dir: Path) -> list[dict]:
+    """Read frames/index.json and return list of {id, path, timestamp}.
+
+    Returns only entries where the image file exists on disk.
+    Paths are resolved relative to *frames_dir* if needed.
+    """
+    index_path = frames_dir / "index.json"
+    if not index_path.exists():
+        return []
+    try:
+        index = json.loads(index_path.read_text())
+    except (json.JSONDecodeError, IOError):
+        return []
+    result = []
+    for entry in index:
+        fpath = Path(entry["path"])
+        if not fpath.exists():
+            fpath = frames_dir / fpath.name
+        if not fpath.exists():
+            continue
+        result.append({
+            "id": entry["id"],
+            "path": fpath,
+            "timestamp": entry.get("timestamp", 0),
+        })
+    return result

cht/stream/lifecycle.py

@@ -0,0 +1,179 @@
+"""Stream lifecycle — manages recording, scene detection, audio extraction, and transcription buffering."""
+
+import logging
+from threading import Thread
+
+from gi.repository import GLib
+
+from cht.config import TRANSCRIBE_MIN_CHUNK_S
+from cht.stream.manager import StreamManager
+from cht.stream.tracker import RecordingTracker
+
+log = logging.getLogger(__name__)
+
+
+class StreamLifecycle:
+    """Owns the streaming process state and coordinates background tasks.
+
+    The window provides UI-facing callbacks; this class handles the
+    process-management side (recorder, tracker, scene detection, audio
+    extraction, transcription buffering).
+    """
+
+    def __init__(self, *, timeline, waveform_engine, transcriber,
+                 on_new_frames, on_waveform_update, on_transcript_ready,
+                 on_scene_marker, on_recorder_restarted):
+        self._timeline = timeline
+        self._waveform_engine = waveform_engine
+        self._transcriber = transcriber
+
+        # Callbacks
+        self._on_new_frames = on_new_frames
+        self._on_waveform_update = on_waveform_update
+        self._on_transcript_ready = on_transcript_ready
+        self._on_scene_marker = on_scene_marker
+        self._on_recorder_restarted = on_recorder_restarted
+
+        # State
+        self._streaming = False
+        self._gone_live = False
+        self._stream_mgr: StreamManager | None = None
+        self._tracker: RecordingTracker | None = None
+        self._pending_transcript_audio: list[tuple] = []
+        self._pending_transcript_duration = 0.0
+
+    @property
+    def is_streaming(self) -> bool:
+        return self._streaming
+
+    @property
+    def stream_mgr(self) -> StreamManager | None:
+        return self._stream_mgr
+
+    @property
+    def tracker(self) -> RecordingTracker | None:
+        return self._tracker
+
+    def start(self, session_id=None) -> StreamManager:
+        """Start recording and all background processes. Returns the StreamManager."""
+        self._streaming = True
+        self._gone_live = False
+
+        self._stream_mgr = StreamManager(session_id=session_id)
+        self._stream_mgr.setup_dirs()
+        self._stream_mgr.start_recorder()
+
+        self._tracker = RecordingTracker(
+            get_segments=lambda: self._stream_mgr.recording_segments if self._stream_mgr else [],
+            on_duration_update=self._on_duration_update,
+        )
+        self._tracker.start()
+
+        self._stream_mgr.start_scene_detector(on_new_frames=self._handle_new_scene_frames)
+        self._stream_mgr.start_audio_extractor(on_new_audio=self._handle_new_audio)
+
+        GLib.timeout_add(1000, self._tick_live)
+        GLib.timeout_add(2000, self._check_recorder)
+
+        return self._stream_mgr
+
+    def stop(self):
+        """Stop all processes and reset state. Does NOT touch UI — caller handles that."""
+        if self._tracker:
+            self._tracker.stop()
+            self._tracker = None
+
+        readonly = self._stream_mgr.readonly if self._stream_mgr else True
+        if self._stream_mgr:
+            if not readonly:
+                self._stream_mgr.stop_all()
+            self._stream_mgr = None
+
+        self._streaming = False
+        self._gone_live = False
+        self._pending_transcript_audio.clear()
+        self._pending_transcript_duration = 0.0
+
+    def set_manager_readonly(self, mgr: StreamManager):
+        """Set a read-only stream manager (for loaded sessions, no streaming)."""
+        self._stream_mgr = mgr
+
+    def clear_manager(self):
+        """Clear the stream manager without stopping processes."""
+        self._stream_mgr = None
+
+    # -- Internal callbacks --
+
+    def _on_duration_update(self, duration):
+        GLib.idle_add(self._timeline.set_duration, duration)
+        if not self._gone_live:
+            self._gone_live = True
+            GLib.idle_add(self._go_live_once)
+            if self._stream_mgr:
+                self._stream_mgr.capture_now(on_new_frames=self._handle_new_scene_frames)
+
+    def _go_live_once(self):
+        if self._stream_mgr:
+            log.info("Going LIVE (startup delay elapsed)")
+            self._timeline.go_live()
+        return False
+
+    def _tick_live(self):
+        if not self._streaming:
+            return False
+        self._timeline.tick_live()
+        return True
+
+    def _handle_new_scene_frames(self, frames):
+        for f in frames:
+            GLib.idle_add(self._on_scene_marker, f["timestamp"])
+        self._on_new_frames(frames)
+
+    def _handle_new_audio(self, wav_path, start_time, duration):
+        if not self._stream_mgr:
+            return
+        self._waveform_engine.append_chunk(wav_path, start_time)
+        peaks = self._waveform_engine.peaks
+        bucket_dur = self._waveform_engine.bucket_duration
+        GLib.idle_add(self._on_waveform_update, peaks.copy(), bucket_dur)
+
+        self._pending_transcript_audio.append((wav_path, start_time, duration))
+        self._pending_transcript_duration += duration
+        if self._pending_transcript_duration < TRANSCRIBE_MIN_CHUNK_S:
+            return
+
+        first_start = self._pending_transcript_audio[0][1]
+        total_dur = self._pending_transcript_duration
+        self._pending_transcript_audio.clear()
+        self._pending_transcript_duration = 0.0
+
+        mgr = self._stream_mgr
+        chunk_wav = mgr.audio_dir / f"transcript_{int(first_start):06d}.wav"
+
+        def _transcribe():
+            from cht.stream import ffmpeg as ff
+            try:
+                ff.extract_audio_chunk(
+                    mgr.recording_path, chunk_wav,
+                    start_time=first_start, duration=total_dur,
+                )
+            except Exception as e:
+                log.error("Transcript audio extraction failed: %s", e)
+                return
+            if not chunk_wav.exists():
+                return
+            new_segs = self._transcriber.transcribe_chunk(chunk_wav, time_offset=first_start)
+            self._transcriber.save_index(mgr.transcript_dir / "index.json")
+            if new_segs:
+                GLib.idle_add(self._on_transcript_ready, new_segs)
+
+        Thread(target=_transcribe, daemon=True, name="transcriber").start()
+
+    def _check_recorder(self):
+        if not self._streaming or not self._stream_mgr:
+            return False
+        if not self._stream_mgr.recorder_alive():
+            log.warning("Recorder died — restarting into new segment")
+            self._stream_mgr.restart_recorder()
+            self._on_recorder_restarted(self._stream_mgr.recording_path)
+        return True

cht/ui/agent_input.py

@@ -0,0 +1,125 @@
+"""Agent input panel — entry, action buttons, model/language dropdowns."""
+
+import logging
+
+import gi
+gi.require_version("Gtk", "4.0")
+from gi.repository import Gtk, GObject
+
+from cht.agent.runner import ACTIONS
+from cht.transcriber.engine import LANGUAGES
+
+log = logging.getLogger(__name__)
+
+
+class AgentInputPanel(Gtk.Frame):
+    """Input bar with action buttons, model/lang selectors, and text entry."""
+
+    __gsignals__ = {
+        "send-requested": (GObject.SignalFlags.RUN_FIRST, None, (str,)),
+        "action-requested": (GObject.SignalFlags.RUN_FIRST, None, (str,)),
+        "model-changed": (GObject.SignalFlags.RUN_FIRST, None, (str,)),
+        "lang-changed": (GObject.SignalFlags.RUN_FIRST, None, (str,)),
+        "history-toggled": (GObject.SignalFlags.RUN_FIRST, None, (bool,)),
+    }
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+        outer = Gtk.Box(orientation=Gtk.Orientation.VERTICAL, spacing=4)
+        outer.set_margin_start(4)
+        outer.set_margin_end(4)
+        outer.set_margin_top(4)
+        outer.set_margin_bottom(4)
+
+        actions_box = Gtk.Box(orientation=Gtk.Orientation.HORIZONTAL, spacing=4)
+        for label, verb in ACTIONS.items():
+            btn = Gtk.Button(label=label)
+            btn.add_css_class("flat")
+            btn.connect("clicked", lambda b, v=verb: self.emit("action-requested", v))
+            actions_box.append(btn)
+
+        spacer = Gtk.Box()
+        spacer.set_hexpand(True)
+        actions_box.append(spacer)
+
+        model_label = Gtk.Label(label="Model:")
+        model_label.add_css_class("dim-label")
+        actions_box.append(model_label)
+
+        self._model_dropdown = Gtk.DropDown.new_from_strings([])
+        self._model_dropdown.set_size_request(200, -1)
+        self._model_dropdown.connect("notify::selected", self._on_model_changed)
+        actions_box.append(self._model_dropdown)
+
+        lang_label = Gtk.Label(label="Lang:")
+        lang_label.add_css_class("dim-label")
+        actions_box.append(lang_label)
+
+        lang_names = list(LANGUAGES.keys())
+        self._lang_names = lang_names
+        self._lang_dropdown = Gtk.DropDown.new_from_strings(lang_names)
+        self._lang_dropdown.set_selected(0)
+        self._lang_dropdown.connect("notify::selected", self._on_lang_changed)
+        actions_box.append(self._lang_dropdown)
+
+        history_toggle = Gtk.CheckButton(label="Chat")
+        history_toggle.set_tooltip_text("Include conversation history in prompts")
+        history_toggle.connect("toggled", lambda b: self.emit("history-toggled", b.get_active()))
+        actions_box.append(history_toggle)
+
+        outer.append(actions_box)
+
+        input_row = Gtk.Box(orientation=Gtk.Orientation.HORIZONTAL, spacing=4)
+        self._entry = Gtk.Entry()
+        self._entry.set_hexpand(True)
+        self._entry.set_placeholder_text("Message agent... (@F1-3 frames, @T1-5 transcript)")
+        self._entry.connect("activate", lambda e: self._do_send())
+        input_row.append(self._entry)
+
+        send_btn = Gtk.Button(label="Send")
+        send_btn.add_css_class("suggested-action")
+        send_btn.connect("clicked", lambda b: self._do_send())
+        input_row.append(send_btn)
+        outer.append(input_row)
+
+        self.set_child(outer)
+
+    @property
+    def entry(self) -> Gtk.Entry:
+        """The text entry widget (for focus checks)."""
+        return self._entry
+
+    def get_text(self) -> str:
+        return self._entry.get_text().strip()
+
+    def clear_text(self) -> None:
+        self._entry.set_text("")
+
+    def populate_models(self, models: list[str], current: str | None = None) -> None:
+        if not models:
+            return
+        string_list = Gtk.StringList.new(models)
+        self._model_dropdown.set_model(string_list)
+        if current:
+            for i, m in enumerate(models):
+                if m == current:
+                    self._model_dropdown.set_selected(i)
+                    break
+
+    def _do_send(self):
+        text = self.get_text()
+        self.clear_text()
+        self.emit("send-requested", text)
+
+    def _on_model_changed(self, dropdown, _pspec):
+        idx = dropdown.get_selected()
+        model = dropdown.get_model()
+        if model and idx < model.get_n_items():
+            self.emit("model-changed", model.get_string(idx))
+
+    def _on_lang_changed(self, dropdown, _pspec):
+        idx = dropdown.get_selected()
+        if idx < len(self._lang_names):
+            lang_code = LANGUAGES[self._lang_names[idx]]
+            self.emit("lang-changed", lang_code or "")

cht/ui/session_dialog.py

@@ -0,0 +1,106 @@
+"""Session browser dialog — lists sessions, supports load and delete."""
+
+import json
+import logging
+
+import gi
+gi.require_version("Gtk", "4.0")
+gi.require_version("Adw", "1")
+from gi.repository import Gtk, Adw, GObject
+
+from cht.stream.manager import list_sessions, delete_sessions
+
+log = logging.getLogger(__name__)
+
+
+class SessionDialog(Adw.Window):
+    """Modal dialog listing sessions. Emits 'session-selected' with session id."""
+
+    __gsignals__ = {
+        "session-selected": (GObject.SignalFlags.RUN_FIRST, None, (str,)),
+    }
+
+    def __init__(self, parent, **kwargs):
+        super().__init__(transient_for=parent, modal=True, **kwargs)
+        self.set_title("Load Session")
+        self.set_default_size(500, 400)
+
+        sessions = list_sessions()
+
+        toolbar = Adw.ToolbarView()
+        header = Adw.HeaderBar()
+
+        select_all_btn = Gtk.CheckButton(label="All")
+        header.pack_start(select_all_btn)
+
+        delete_btn = Gtk.Button(label="Delete")
+        delete_btn.add_css_class("destructive-action")
+        header.pack_end(delete_btn)
+
+        toolbar.add_top_bar(header)
+
+        scroll = Gtk.ScrolledWindow()
+        scroll.set_vexpand(True)
+        listbox = Gtk.ListBox()
+        listbox.set_selection_mode(Gtk.SelectionMode.NONE)
+        listbox.add_css_class("boxed-list")
+
+        checks: list[tuple[str, Gtk.CheckButton]] = []
+
+        for sid, sdir in sessions:
+            idx = sdir / "frames" / "index.json"
+            nframes = 0
+            try:
+                nframes = len(json.loads(idx.read_text()))
+            except Exception:
+                pass
+            nrec = len(list((sdir / "stream").glob("recording_*.mp4")))
+
+            check = Gtk.CheckButton()
+            checks.append((sid, check))
+
+            row = Adw.ActionRow()
+            row.set_title(sid)
+            row.set_subtitle(f"{nframes} frames, {nrec} segments")
+            row.set_activatable(True)
+            row.add_prefix(check)
+
+            def _on_row_activated(r, s=sid):
+                self.close()
+                self.emit("session-selected", s)
+            row.connect("activated", _on_row_activated)
+            listbox.append(row)
+
+        def _on_select_all(btn):
+            active = btn.get_active()
+            for _, cb in checks:
+                cb.set_active(active)
+        select_all_btn.connect("toggled", _on_select_all)
+
+        def _on_delete(btn):
+            to_delete = [sid for sid, cb in checks if cb.get_active()]
+            if not to_delete:
+                return
+            if self._current_session in to_delete:
+                to_delete.remove(self._current_session)
+            if to_delete:
+                delete_sessions(to_delete)
+                self.close()
+                # Re-open with updated list
+                new_dialog = SessionDialog(self.get_transient_for())
+                new_dialog._current_session = self._current_session
+                # Forward the signal
+                new_dialog.connect("session-selected",
+                                   lambda d, s: self.emit("session-selected", s))
+                new_dialog.present()
+        delete_btn.connect("clicked", _on_delete)
+
+        scroll.set_child(listbox)
+        toolbar.set_content(scroll)
+        self.set_content(toolbar)
+
+        self._current_session = None
+
+    def set_current_session(self, session_id: str | None) -> None:
+        """Set the active session id so it won't be deleted."""
+        self._current_session = session_id

cht/window.py

@@ -1,18 +1,16 @@
 """Main application window — wires Timeline to all components."""
 
-import json
 import logging
-from pathlib import Path
 
 import gi
 gi.require_version("Gtk", "4.0")
 gi.require_version("Adw", "1")
 gi.require_version("GdkPixbuf", "2.0")
-from gi.repository import Gtk, Gdk, Adw, GLib, Pango, GdkPixbuf
+from gi.repository import Gtk, Gdk, Adw, GLib, GdkPixbuf
 
 from threading import Thread
 
-from cht.config import APP_NAME, SCENE_THRESHOLD, TRANSCRIBE_MIN_CHUNK_S
+from cht.config import APP_NAME
 from cht.ui.timeline import Timeline, TimelineControls
 from cht.ui.monitor import MonitorWidget
 from cht.ui.waveform import WaveformWidget
@@ -20,11 +18,14 @@ from cht.ui.frames_panel import FramesPanel
 from cht.ui.transcript_panel import TranscriptPanel
 from cht.ui.keyboard import KeyboardManager, KEY_LEFT, KEY_RIGHT, KEY_UP, KEY_DOWN, KEY_RETURN, KEY_KP_ENTER, KEY_ESCAPE, KEY_DELETE
 from cht.ui.agent_output import AgentOutputPanel
+from cht.ui.agent_input import AgentInputPanel
 from cht.audio.waveform import WaveformEngine
-from cht.transcriber.engine import TranscriberEngine, LANGUAGES
-from cht.stream.manager import StreamManager, list_sessions, delete_sessions
-from cht.stream.tracker import RecordingTracker
-from cht.agent.runner import AgentRunner, ACTIONS, check_claude_cli
+from cht.transcriber.engine import TranscriberEngine
+from cht.stream.manager import StreamManager, list_sessions
+from cht.stream.lifecycle import StreamLifecycle
+from cht.ui.session_dialog import SessionDialog
+from cht.session import load_frame_index
+from cht.agent.runner import AgentRunner, check_claude_cli
 
 log = logging.getLogger(__name__)
 
@@ -34,10 +35,6 @@ class ChtWindow(Adw.ApplicationWindow):
         super().__init__(**kwargs)
         self.set_title(APP_NAME)
         self.set_default_size(1400, 900)
-        self._streaming = False
-        self._gone_live = False
-        self._stream_mgr = None
-        self._tracker = None
         self._known_frames = set()
 
         # Core components
@@ -45,8 +42,19 @@ class ChtWindow(Adw.ApplicationWindow):
         self._agent = AgentRunner()
         self._waveform_engine = WaveformEngine()
         self._transcriber = TranscriberEngine()
-        self._pending_transcript_audio = []
-        self._pending_transcript_duration = 0.0
+
+        # Stream lifecycle (owns streaming state, recorder, tracker, audio buffering)
+        # Lambdas used because panels/widgets aren't created yet at this point.
+        self._lifecycle = StreamLifecycle(
+            timeline=self._timeline,
+            waveform_engine=self._waveform_engine,
+            transcriber=self._transcriber,
+            on_new_frames=lambda frames: None,  # frame polling handles new frames
+            on_waveform_update=lambda peaks, bd: self._waveform_widget.set_peaks(peaks, bd),
+            on_transcript_ready=lambda segs: self._transcript_panel.add_items(segs),
+            on_scene_marker=lambda ts: self._timeline.add_scene_marker(ts),
+            on_recorder_restarted=lambda path: self._monitor.set_recording(path),
+        )
 
         # Panels (own their selection state)
         self._frames_panel = FramesPanel()
@@ -112,21 +120,23 @@ class ChtWindow(Adw.ApplicationWindow):
     # -- Connect / Disconnect --
 
     def _on_connect_clicked(self, button):
-        if self._streaming:
+        if self._lifecycle.is_streaming:
             self._stop_stream(reload_session=True)
         else:
-            session_id = self._stream_mgr.session_id if self._stream_mgr else None
-            if self._stream_mgr:
+            session_id = self._lifecycle.stream_mgr.session_id if self._lifecycle.stream_mgr else None
+            if self._lifecycle.stream_mgr:
                 self._stop_stream()
             self._start_stream(session_id=session_id)
 
     def _on_capture_clicked(self):
-        if self._stream_mgr:
-            self._stream_mgr.capture_now(on_new_frames=self._on_new_scene_frames)
+        if self._lifecycle.stream_mgr:
+            self._lifecycle.stream_mgr.capture_now(
+                on_new_frames=self._lifecycle._handle_new_scene_frames
+            )
 
     def _on_scene_threshold(self, val):
-        if self._stream_mgr:
-            self._stream_mgr.scene_threshold = val
+        if self._lifecycle.stream_mgr:
+            self._lifecycle.stream_mgr.scene_threshold = val
 
     def _on_min_chunk_changed(self, panel, val):
         import cht.config
@@ -143,97 +153,33 @@ class ChtWindow(Adw.ApplicationWindow):
         if not sessions:
             self._agent_output.append("No previous sessions found.\n")
             return
-
-        dialog = Adw.Window(transient_for=self, modal=True)
-        dialog.set_title("Load Session")
-        dialog.set_default_size(500, 400)
-
-        toolbar = Adw.ToolbarView()
-        header = Adw.HeaderBar()
-
-        select_all_btn = Gtk.CheckButton(label="All")
-        header.pack_start(select_all_btn)
-
-        delete_btn = Gtk.Button(label="Delete")
-        delete_btn.add_css_class("destructive-action")
-        header.pack_end(delete_btn)
-
-        toolbar.add_top_bar(header)
-
-        scroll = Gtk.ScrolledWindow()
-        scroll.set_vexpand(True)
-        listbox = Gtk.ListBox()
-        listbox.set_selection_mode(Gtk.SelectionMode.NONE)
-        listbox.add_css_class("boxed-list")
-
-        checks: list[tuple[str, Gtk.CheckButton]] = []
-
-        for sid, sdir in sessions:
-            idx = sdir / "frames" / "index.json"
-            nframes = 0
-            try:
-                nframes = len(json.loads(idx.read_text()))
-            except Exception:
-                pass
-            nrec = len(list((sdir / "stream").glob("recording_*.mp4")))
-
-            check = Gtk.CheckButton()
-            checks.append((sid, check))
-
-            row = Adw.ActionRow()
-            row.set_title(sid)
-            row.set_subtitle(f"{nframes} frames, {nrec} segments")
-            row.set_activatable(True)
-            row.add_prefix(check)
-
-            def _on_row_activated(r, s=sid, d=dialog):
-                d.close()
-                self._load_session(s)
-            row.connect("activated", _on_row_activated)
-            listbox.append(row)
-
-        def _on_select_all(btn):
-            active = btn.get_active()
-            for _, cb in checks:
-                cb.set_active(active)
-        select_all_btn.connect("toggled", _on_select_all)
-
-        def _on_delete(btn):
-            to_delete = [sid for sid, cb in checks if cb.get_active()]
-            if not to_delete:
-                return
-            current = self._stream_mgr.session_id if self._stream_mgr else None
-            if current in to_delete:
-                to_delete.remove(current)
-            if to_delete:
-                delete_sessions(to_delete)
-                dialog.close()
-                self._on_load_session_clicked(None)
-        delete_btn.connect("clicked", _on_delete)
-
-        scroll.set_child(listbox)
-        toolbar.set_content(scroll)
-        dialog.set_content(toolbar)
+        dialog = SessionDialog(self)
+        dialog.set_current_session(
+            self._lifecycle.stream_mgr.session_id if self._lifecycle.stream_mgr else None
+        )
+        dialog.connect("session-selected", lambda d, sid: self._load_session(sid))
         dialog.present()
 
     def _load_session(self, session_id):
         """Load an existing session for review (no streaming)."""
-        if self._streaming or self._stream_mgr:
+        if self._lifecycle.is_streaming or self._lifecycle.stream_mgr:
             self._stop_stream()
 
         try:
-            self._stream_mgr = StreamManager.from_existing(session_id)
+            mgr = StreamManager.from_existing(session_id)
         except FileNotFoundError as e:
             self._agent_output.append(f"Error: {e}\n")
             return
 
+        self._lifecycle.set_manager_readonly(mgr)
+
         self.set_title(f"{APP_NAME} — {session_id}")
         self._agent_output.append(f"Loaded session: {session_id}\n")
 
-        segments = self._stream_mgr.recording_segments
+        segments = mgr.recording_segments
         if segments:
             self._monitor.set_recording(segments[0])
-            duration = self._stream_mgr.total_duration()
+            duration = mgr.total_duration()
             if duration > 0:
                 self._timeline.set_duration(duration)
                 self._timeline.seek(0)
@@ -252,7 +198,7 @@ class ChtWindow(Adw.ApplicationWindow):
             from cht.stream import ffmpeg as ff
 
             def _compute_waveform():
-                audio_dir = self._stream_mgr.audio_dir
+                audio_dir = mgr.audio_dir
                 audio_dir.mkdir(parents=True, exist_ok=True)
                 full_wav = audio_dir / "full.wav"
                 try:
@@ -275,125 +221,32 @@ class ChtWindow(Adw.ApplicationWindow):
         self._connect_btn.set_label("Disconnect")
         self._connect_btn.remove_css_class("suggested-action")
         self._connect_btn.add_css_class("destructive-action")
-        self._streaming = True
-        self._gone_live = False
 
-        self._stream_mgr = StreamManager(session_id=session_id)
-        self._stream_mgr.setup_dirs()
-        self._stream_mgr.start_recorder()
+        mgr = self._lifecycle.start(session_id=session_id)
 
-        self._monitor.set_recording(self._stream_mgr.recording_path)
-        self._monitor.set_live_source(self._stream_mgr.relay_url)
-
-        self._tracker = RecordingTracker(
-            get_segments=lambda: self._stream_mgr.recording_segments if self._stream_mgr else [],
-            on_duration_update=self._on_duration_update,
-        )
-        self._tracker.start()
-
-        self._stream_mgr.start_scene_detector(on_new_frames=self._on_new_scene_frames)
-        self._stream_mgr.start_audio_extractor(on_new_audio=self._on_new_audio)
+        self._monitor.set_recording(mgr.recording_path)
+        self._monitor.set_live_source(mgr.relay_url)
 
         GLib.timeout_add(1000, self._poll_frames)
-        GLib.timeout_add(1000, self._tick_live)
-        GLib.timeout_add(2000, self._check_recorder)
 
         # Reload existing data if resuming
         if session_id:
             self._load_existing_frames()
             self._load_existing_transcript()
 
-        self.set_title(f"{APP_NAME} — {self._stream_mgr.session_id}")
+        self.set_title(f"{APP_NAME} — {mgr.session_id}")
         log.info("Waiting for sender...")
 
-    def _go_live_once(self):
-        if self._stream_mgr:
-            log.info("Going LIVE (startup delay elapsed)")
-            self._timeline.go_live()
-        return False
-
-    def _tick_live(self):
-        if not self._streaming:
-            return False
-        self._timeline.tick_live()
-        return True
-
-    def _on_duration_update(self, duration):
-        GLib.idle_add(self._timeline.set_duration, duration)
-        if not self._gone_live:
-            self._gone_live = True
-            GLib.idle_add(self._go_live_once)
-            if self._stream_mgr:
-                self._stream_mgr.capture_now(on_new_frames=self._on_new_scene_frames)
-
-    def _on_new_scene_frames(self, frames):
-        for f in frames:
-            GLib.idle_add(self._timeline.add_scene_marker, f["timestamp"])
-
-    def _on_new_audio(self, wav_path, start_time, duration):
-        if not self._stream_mgr:
-            return
-        self._waveform_engine.append_chunk(wav_path, start_time)
-        peaks = self._waveform_engine.peaks
-        bucket_dur = self._waveform_engine.bucket_duration
-        GLib.idle_add(self._waveform_widget.set_peaks, peaks.copy(), bucket_dur)
-
-        self._pending_transcript_audio.append((wav_path, start_time, duration))
-        self._pending_transcript_duration += duration
-        if self._pending_transcript_duration < TRANSCRIBE_MIN_CHUNK_S:
-            return
-
-        first_start = self._pending_transcript_audio[0][1]
-        total_dur = self._pending_transcript_duration
-        self._pending_transcript_audio.clear()
-        self._pending_transcript_duration = 0.0
-
-        mgr = self._stream_mgr
-        chunk_wav = mgr.audio_dir / f"transcript_{int(first_start):06d}.wav"
-
-        def _transcribe():
-            from cht.stream import ffmpeg as ff
-            try:
-                ff.extract_audio_chunk(
-                    mgr.recording_path, chunk_wav,
-                    start_time=first_start, duration=total_dur,
-                )
-            except Exception as e:
-                log.error("Transcript audio extraction failed: %s", e)
-                return
-            if not chunk_wav.exists():
-                return
-            new_segs = self._transcriber.transcribe_chunk(chunk_wav, time_offset=first_start)
-            self._transcriber.save_index(mgr.transcript_dir / "index.json")
-            if new_segs:
-                GLib.idle_add(self._transcript_panel.add_items, new_segs)
-
-        Thread(target=_transcribe, daemon=True, name="transcriber").start()
-
-    def _check_recorder(self):
-        if not self._streaming or not self._stream_mgr:
-            return False
-        if not self._stream_mgr.recorder_alive():
-            log.warning("Recorder died — restarting into new segment")
-            self._stream_mgr.restart_recorder()
-            self._monitor.set_recording(self._stream_mgr.recording_path)
-        return True
-
     def _on_live_toggle(self):
         pos = self._monitor.get_live_position()
         self._timeline.toggle_live(live_player_pos=pos)
 
     def _stop_stream(self, reload_session=False):
         log.info("Stopping stream...")
-        last_session_id = self._stream_mgr.session_id if self._stream_mgr and not self._stream_mgr.readonly else None
+        mgr = self._lifecycle.stream_mgr
+        last_session_id = mgr.session_id if mgr and not mgr.readonly else None
 
-        if self._tracker:
-            self._tracker.stop()
-            self._tracker = None
-        if self._stream_mgr:
-            if not self._stream_mgr.readonly:
-                self._stream_mgr.stop_all()
-            self._stream_mgr = None
+        self._lifecycle.stop()
 
         self._timeline.reset()
         self._monitor.reset()
@@ -401,8 +254,6 @@ class ChtWindow(Adw.ApplicationWindow):
         self._waveform_widget.set_peaks(None, 0.05)
         self._transcriber.reset()
         self._agent.clear_history()
-        self._pending_transcript_audio.clear()
-        self._pending_transcript_duration = 0.0
         self._known_frames = set()
 
         self._frames_panel.clear()
@@ -411,7 +262,6 @@ class ChtWindow(Adw.ApplicationWindow):
         self._connect_btn.set_label("Connect")
         self._connect_btn.remove_css_class("destructive-action")
         self._connect_btn.add_css_class("suggested-action")
-        self._streaming = False
         self.set_title(APP_NAME)
 
         if reload_session and last_session_id:
@@ -422,7 +272,7 @@ class ChtWindow(Adw.ApplicationWindow):
 
     def teardown(self):
         """Full cleanup for app exit — safe to call multiple times."""
-        if self._stream_mgr or self._streaming:
+        if self._lifecycle.stream_mgr or self._lifecycle.is_streaming:
             self._stop_stream()
         self._monitor.stop()
 
@@ -465,72 +315,16 @@ class ChtWindow(Adw.ApplicationWindow):
         right_box.append(transcript_frame)
 
         # Agent input
-        self._agent_input = self._build_agent_input()
+        self._agent_input = AgentInputPanel()
+        self._agent_input.connect("send-requested", lambda p, text: self._send_message(text or None))
+        self._agent_input.connect("action-requested", lambda p, verb: self._send_action(verb))
+        self._agent_input.connect("model-changed", self._on_model_changed)
+        self._agent_input.connect("lang-changed", self._on_lang_changed)
+        self._agent_input.connect("history-toggled", lambda p, v: setattr(self._agent, "include_history", v))
         right_box.append(self._agent_input)
 
         return right_box
 
-    def _build_agent_input(self):
-        outer = Gtk.Box(orientation=Gtk.Orientation.VERTICAL, spacing=4)
-        outer.set_margin_start(4)
-        outer.set_margin_end(4)
-        outer.set_margin_top(4)
-        outer.set_margin_bottom(4)
-
-        actions_box = Gtk.Box(orientation=Gtk.Orientation.HORIZONTAL, spacing=4)
-        for label, verb in ACTIONS.items():
-            btn = Gtk.Button(label=label)
-            btn.add_css_class("flat")
-            btn.connect("clicked", lambda b, v=verb: self._send_action(v))
-            actions_box.append(btn)
-
-        spacer = Gtk.Box()
-        spacer.set_hexpand(True)
-        actions_box.append(spacer)
-
-        model_label = Gtk.Label(label="Model:")
-        model_label.add_css_class("dim-label")
-        actions_box.append(model_label)
-
-        self._model_dropdown = Gtk.DropDown.new_from_strings([])
-        self._model_dropdown.set_size_request(200, -1)
-        self._model_dropdown.connect("notify::selected", self._on_model_changed)
-        actions_box.append(self._model_dropdown)
-
-        lang_label = Gtk.Label(label="Lang:")
-        lang_label.add_css_class("dim-label")
-        actions_box.append(lang_label)
-
-        lang_names = list(LANGUAGES.keys())
-        self._lang_dropdown = Gtk.DropDown.new_from_strings(lang_names)
-        self._lang_dropdown.set_selected(0)
-        self._lang_dropdown.connect("notify::selected", self._on_lang_changed)
-        actions_box.append(self._lang_dropdown)
-
-        self._history_toggle = Gtk.CheckButton(label="Chat")
-        self._history_toggle.set_tooltip_text("Include conversation history in prompts")
-        self._history_toggle.connect("toggled", lambda b: setattr(self._agent, "include_history", b.get_active()))
-        actions_box.append(self._history_toggle)
-
-        outer.append(actions_box)
-
-        input_row = Gtk.Box(orientation=Gtk.Orientation.HORIZONTAL, spacing=4)
-        self._input_entry = Gtk.Entry()
-        self._input_entry.set_hexpand(True)
-        self._input_entry.set_placeholder_text("Message agent... (@F1-3 frames, @T1-5 transcript)")
-        self._input_entry.connect("activate", lambda e: self._send_message())
-        input_row.append(self._input_entry)
-
-        send_btn = Gtk.Button(label="Send")
-        send_btn.add_css_class("suggested-action")
-        send_btn.connect("clicked", lambda b: self._send_message())
-        input_row.append(send_btn)
-        outer.append(input_row)
-
-        frame = Gtk.Frame()
-        frame.set_child(outer)
-        return frame
-
     # -- Keyboard --
 
     def _setup_keyboard(self):
@@ -541,7 +335,7 @@ class ChtWindow(Adw.ApplicationWindow):
                 return False
             w = focus
             while w is not None:
-                if w is self._input_entry:
+                if w is self._agent_input.entry:
                     return True
                 w = w.get_parent()
             return False
@@ -575,14 +369,11 @@ class ChtWindow(Adw.ApplicationWindow):
         self._send_message(msg)
 
     def _send_message(self, text: str | None = None):
-        if text is None:
-            text = self._input_entry.get_text().strip()
-            self._input_entry.set_text("")
         if not text:
             text = self._build_selection_message("answer")
         if not text:
             return
-        if not self._stream_mgr:
+        if not self._lifecycle.stream_mgr:
             self._agent_output.append("No active session.\n")
             return
 
@@ -591,40 +382,26 @@ class ChtWindow(Adw.ApplicationWindow):
         self._agent_output.begin_response()
         self._agent.send(
             message=text,
-            stream_mgr=self._stream_mgr,
-            tracker=self._tracker,
+            stream_mgr=self._lifecycle.stream_mgr,
+            tracker=self._lifecycle.tracker,
             on_chunk=lambda chunk: GLib.idle_add(self._agent_output.replace_thinking, chunk),
             on_done=lambda err: GLib.idle_add(self._agent_output.finish_response, err),
         )
 
     # -- Settings callbacks --
 
-    def _on_lang_changed(self, dropdown, _pspec):
-        idx = dropdown.get_selected()
-        lang_names = list(LANGUAGES.keys())
-        if idx < len(lang_names):
-            lang_code = LANGUAGES[lang_names[idx]]
-            self._transcriber.language = lang_code
-            log.info("Transcript language: %s (%s)", lang_names[idx], lang_code or "auto")
+    def _on_lang_changed(self, _panel, lang_code):
+        self._transcriber.language = lang_code or None
+        log.info("Transcript language: %s", lang_code or "auto")
 
-    def _on_model_changed(self, dropdown, _pspec):
-        idx = dropdown.get_selected()
-        model = self._agent.available_models[idx] if idx < len(self._agent.available_models) else None
-        if model:
-            self._agent.model = model
-            log.info("Model switched to %s", model)
+    def _on_model_changed(self, _panel, model):
+        self._agent.model = model
+        log.info("Model switched to %s", model)
 
     def _populate_model_dropdown(self):
-        models = self._agent.available_models
-        if not models:
-            return
-        string_list = Gtk.StringList.new(models)
-        self._model_dropdown.set_model(string_list)
-        current = self._agent.model
-        for i, m in enumerate(models):
-            if m == current:
-                self._model_dropdown.set_selected(i)
-                break
+        self._agent_input.populate_models(
+            self._agent.available_models, self._agent.model
+        )
 
     def _check_agent_auth(self):
         import os
@@ -641,26 +418,17 @@ class ChtWindow(Adw.ApplicationWindow):
     # -- Data loading --
 
     def _load_existing_frames(self):
-        if not self._stream_mgr:
+        if not self._lifecycle.stream_mgr:
             return
-        index_path = self._stream_mgr.frames_dir / "index.json"
-        if not index_path.exists():
+        entries = load_frame_index(self._lifecycle.stream_mgr.frames_dir)
+        if not entries:
             self._agent_output.append("  No frames found.\n")
             return
-        try:
-            index = json.loads(index_path.read_text())
-        except (json.JSONDecodeError, IOError):
-            return
         items = []
-        for entry in index:
-            fpath = Path(entry["path"])
-            if not fpath.exists():
-                fpath = self._stream_mgr.frames_dir / fpath.name
-            if not fpath.exists():
-                continue
+        for entry in entries:
             try:
-                pixbuf = GdkPixbuf.Pixbuf.new_from_file_at_scale(str(fpath), 256, 144, True)
-                items.append({"id": entry["id"], "pixbuf": pixbuf, "timestamp": entry.get("timestamp", 0)})
+                pixbuf = GdkPixbuf.Pixbuf.new_from_file_at_scale(str(entry["path"]), 256, 144, True)
+                items.append({"id": entry["id"], "pixbuf": pixbuf, "timestamp": entry["timestamp"]})
             except Exception as e:
                 log.warning("Thumbnail load failed for %s: %s", entry["id"], e)
         if items:
@@ -669,9 +437,9 @@ class ChtWindow(Adw.ApplicationWindow):
             self._agent_output.append(f"  Loaded {len(items)} frame thumbnails.\n")
 
     def _load_existing_transcript(self):
-        if not self._stream_mgr:
+        if not self._lifecycle.stream_mgr:
             return
-        transcript_index = self._stream_mgr.transcript_dir / "index.json"
+        transcript_index = self._lifecycle.stream_mgr.transcript_dir / "index.json"
         if not transcript_index.exists():
             return
         self._transcriber.load_index(transcript_index)
@@ -681,28 +449,17 @@ class ChtWindow(Adw.ApplicationWindow):
             self._agent_output.append(f"  Loaded {len(segs)} transcript segments.\n")
 
     def _poll_frames(self):
-        if not self._stream_mgr:
+        if not self._lifecycle.stream_mgr:
             return False
-        index_path = self._stream_mgr.frames_dir / "index.json"
-        if not index_path.exists():
-            return True
-        try:
-            index = json.loads(index_path.read_text())
-        except (json.JSONDecodeError, IOError):
-            return True
-        for entry in index:
+        for entry in load_frame_index(self._lifecycle.stream_mgr.frames_dir):
             fid = entry["id"]
             if fid in self._known_frames:
                 continue
-            fpath = Path(entry["path"])
-            if not fpath.exists():
-                continue
             self._known_frames.add(fid)
-            timestamp = entry.get("timestamp", 0)
             try:
-                pixbuf = GdkPixbuf.Pixbuf.new_from_file_at_scale(str(fpath), 256, 144, True)
+                pixbuf = GdkPixbuf.Pixbuf.new_from_file_at_scale(str(entry["path"]), 256, 144, True)
                 auto = not self._transcript_panel.has_selection
-                self._frames_panel.add_item(fid, pixbuf, timestamp, auto_select=auto)
+                self._frames_panel.add_item(fid, pixbuf, entry["timestamp"], auto_select=auto)
             except Exception as e:
                 log.warning("Thumbnail load failed for %s: %s", fid, e)
         return True