better agent

cht/window.py

@@ -28,6 +28,8 @@ from cht.ui.session_dialog import SessionDialog
 from cht.session import load_frame_index, load_segment_manifest, rebuild_manifest, global_time_to_segment
 from cht.scrub.manager import ProxyManager
 from cht.agent.runner import AgentRunner, check_claude_cli
+from cht.agent.base import TextDelta, ToolCallStart, ToolCallEnd, ToolResult, ToolUse, Done, Error
+from cht.agent.buffer import StreamingTextBuffer
 from cht.telemetry import Telemetry
 
 log = logging.getLogger(__name__)
@@ -245,6 +247,11 @@ class ChtWindow(Adw.ApplicationWindow):
         self._update_scrub_bar_manifest()
         self._populate_model_dropdown()
 
+        # Load persisted agent conversation
+        self._agent.load_from_session(mgr.session_dir)
+        if self._agent.thread.messages:
+            self._agent_output.load_thread(self._agent.thread)
+
     def _reload_waveform(self, mgr):
         """Recompute waveform from existing segments in background."""
         segments = mgr.recording_segments
@@ -289,6 +296,9 @@ class ChtWindow(Adw.ApplicationWindow):
             self._load_existing_frames()
             self._load_existing_transcript()
             self._reload_waveform(mgr)
+            self._agent.load_from_session(mgr.session_dir)
+            if self._agent.thread.messages:
+                self._agent_output.load_thread(self._agent.thread)
 
         self.set_title(f"{APP_NAME} — {mgr.session_id}")
         log.info("Waiting for sender...")
@@ -453,6 +463,11 @@ class ChtWindow(Adw.ApplicationWindow):
         mgr = self._lifecycle.stream_mgr
         last_session_id = mgr.session_id if mgr and not mgr.readonly else None
 
+        # Save agent thread before stopping
+        if mgr and self._agent.thread.messages:
+            from cht.agent.base import save_thread
+            save_thread(self._agent.thread, mgr.session_dir)
+
         if self._telemetry:
             self._telemetry.close()
             self._telemetry = None
@@ -604,15 +619,53 @@ class ChtWindow(Adw.ApplicationWindow):
             self._agent_output.append("No active session.\n")
             return
 
-        self._agent_output.append(f"\n> {text}\n…\n")
+        # Show user message in UI
+        from cht.agent.runner import _parse_mentions, _parse_transcript_mentions
+        from cht.agent.tools import load_frames, load_transcript
+        mgr = self._lifecycle.stream_mgr
+        frames = load_frames(mgr.frames_dir)
+        mentioned_frames = _parse_mentions(text, frames)
+        transcript = load_transcript(mgr.transcript_dir)
+        mentioned_transcripts = _parse_transcript_mentions(text, transcript)
+        self._agent_output.add_user_message(text, mentioned_frames, mentioned_transcripts)
+
+        # Prepare streaming
+        from cht.agent.base import _msg_id
+        msg_id = _msg_id()
+        self._agent_output.begin_assistant_message(msg_id)
+
+        full_text_parts = []
+        buffer = StreamingTextBuffer(
+            on_reveal=lambda chunk: self._agent_output.append_to_assistant(msg_id, chunk)
+        )
+
+        def on_event(event):
+            if isinstance(event, TextDelta):
+                full_text_parts.append(event.text)
+                GLib.idle_add(buffer.push, event.text)
+            elif isinstance(event, ToolCallStart):
+                tu = ToolUse(id=event.id, tool_name=event.name, input=event.input, status="running")
+                GLib.idle_add(self._agent_output.add_tool_call, tu)
+            elif isinstance(event, ToolResult):
+                GLib.idle_add(self._agent_output.update_tool_result, event.tool_use_id, event)
+            elif isinstance(event, Error):
+                GLib.idle_add(self._agent_output.append, f"[Error: {event.message}]\n")
+
+        def on_done(err):
+            def _finish():
+                buffer.flush()
+                if err:
+                    self._agent_output.append(f"[Error: {err}]\n")
+                else:
+                    self._agent_output.finish_assistant(msg_id, "".join(full_text_parts))
+            GLib.idle_add(_finish)
 
-        self._agent_output.begin_response()
         self._agent.send(
             message=text,
-            stream_mgr=self._lifecycle.stream_mgr,
+            stream_mgr=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),
+            on_event=on_event,
+            on_done=on_done,
         )
 
     # -- Settings callbacks --