diff --git a/cht/agent/claude_sdk_provider.py b/cht/agent/claude_sdk_provider.py
index ea5b42a..1078860 100644
--- a/cht/agent/claude_sdk_provider.py
+++ b/cht/agent/claude_sdk_provider.py
@@ -27,36 +27,18 @@ Be concise and specific. Focus on what's visible in the frames."""
 
 def _build_prompt(message: str, context: SessionContext) -> str:
     lines = []
-
     # Session summary
     m, s = divmod(int(context.duration), 60)
     lines.append(f"Recording duration: {m:02d}:{s:02d}")
     lines.append(f"Total frames captured: {len(context.frames)}")
 
-    # All available frames (let Claude decide which to look at)
-    if context.frames:
-        lines.append("\nAvailable frames:")
-        for f in context.frames:
-            fm, fs = divmod(int(f.timestamp), 60)
-            lines.append(f"  {f.id} at {fm:02d}:{fs:02d} — {f.path}")
-
-    # Explicitly mentioned frames
     if context.mentioned_frames:
-        lines.append("\nFrames referenced in this message:")
+        lines.append("\nFrames:")
         for f in context.mentioned_frames:
             fm, fs = divmod(int(f.timestamp), 60)
             lines.append(f"  {f.id} at {fm:02d}:{fs:02d} — {f.path}")
-
-    # Transcript
-    if context.transcript_segments:
-        lines.append(f"\nTranscript ({len(context.transcript_segments)} segments):")
-        for t in context.transcript_segments:
-            tm1, ts1 = divmod(int(t.start), 60)
-            tm2, ts2 = divmod(int(t.end), 60)
-            lines.append(f"  {t.id} [{tm1:02d}:{ts1:02d}-{tm2:02d}:{ts2:02d}] {t.text}")
-
     if context.mentioned_transcripts:
-        lines.append("\nTranscript segments referenced in this message:")
+        lines.append("\nTranscript:")
         for t in context.mentioned_transcripts:
             tm1, ts1 = divmod(int(t.start), 60)
             tm2, ts2 = divmod(int(t.end), 60)
diff --git a/cht/agent/openai_compat_provider.py b/cht/agent/openai_compat_provider.py
index 14eff50..89fb6ac 100644
--- a/cht/agent/openai_compat_provider.py
+++ b/cht/agent/openai_compat_provider.py
@@ -99,9 +99,9 @@ class OpenAICompatProvider(AgentProvider):
             f"Recording duration: {m:02d}:{s:02d}",
             f"Total frames: {len(context.frames)}",
         ]
-        if context.transcript_segments:
-            ctx_lines.append(f"\nTranscript ({len(context.transcript_segments)} segments):")
-            for t in context.transcript_segments:
+        if context.mentioned_transcripts:
+            ctx_lines.append("\nTranscript:")
+            for t in context.mentioned_transcripts:
                 tm1, ts1 = divmod(int(t.start), 60)
                 tm2, ts2 = divmod(int(t.end), 60)
                 ctx_lines.append(f"  {t.id} [{tm1:02d}:{ts1:02d}-{tm2:02d}:{ts2:02d}] {t.text}")
diff --git a/cht/ui/keyboard.py b/cht/ui/keyboard.py
index e368683..6909a4e 100644
--- a/cht/ui/keyboard.py
+++ b/cht/ui/keyboard.py
@@ -41,15 +41,18 @@ class KeyboardManager:
     def __init__(self):
         self._bindings: dict[int, Callable] = {}
         self._passthrough: Callable[[], bool] | None = None
+        self._passthrough_except: set[int] = set()
         self._window = None
 
     def bind(self, keyval: int, handler: Callable):
         """Register a handler for a key. Handler receives shift=bool."""
         self._bindings[keyval] = handler
 
-    def set_passthrough(self, check: Callable[[], bool]):
-        """When check() returns True, keys pass through to focused widget."""
+    def set_passthrough(self, check: Callable[[], bool], except_keys: set[int] | None = None):
+        """When check() returns True, keys pass through to focused widget.
+        Keys in except_keys are still handled even during passthrough."""
         self._passthrough = check
+        self._passthrough_except = except_keys or set()
 
     def attach(self, window):
         """Attach to a GTK4 window."""
@@ -76,7 +79,7 @@ class KeyboardManager:
             self._window.set_focus(None)
 
     def _on_key_pressed(self, controller, keyval, keycode, state):
-        if self._passthrough and self._passthrough():
+        if self._passthrough and self._passthrough() and keyval not in self._passthrough_except:
             return False
 
         handler = self._bindings.get(keyval)
diff --git a/cht/window.py b/cht/window.py
index 6b5e447..20c5be3 100644
--- a/cht/window.py
+++ b/cht/window.py
@@ -563,14 +563,14 @@ class ChtWindow(Adw.ApplicationWindow):
 
     def _setup_keyboard(self):
         kb = KeyboardManager()
-        kb.set_passthrough(lambda: self.get_focus() is self._input_entry)
+        kb.set_passthrough(lambda: self.get_focus() is self._input_entry, except_keys={KEY_ESCAPE})
         kb.bind(KEY_LEFT, lambda **_: self._frames_panel.select_adjacent(-1))
         kb.bind(KEY_RIGHT, lambda **_: self._frames_panel.select_adjacent(1))
         kb.bind(KEY_UP, lambda shift=False, **_: self._transcript_panel.select_adjacent(-1, extend=shift))
         kb.bind(KEY_DOWN, lambda shift=False, **_: self._transcript_panel.select_adjacent(1, extend=shift))
         kb.bind(KEY_RETURN, lambda **_: self._send_message(self._build_selection_message("answer")) if self._build_selection_message("answer") else None)
         kb.bind(KEY_KP_ENTER, lambda **_: self._send_message(self._build_selection_message("answer")) if self._build_selection_message("answer") else None)
-        kb.bind(KEY_ESCAPE, lambda **_: (self._frames_panel.clear_selection(), self._transcript_panel.clear_selection()))
+        kb.bind(KEY_ESCAPE, lambda **_: (self.set_focus(None), self._frames_panel.clear_selection(), self._transcript_panel.clear_selection()))
         kb.bind(KEY_DELETE, lambda **_: self._on_clear_agent_output(None))
         kb.attach(self)
 
@@ -596,6 +596,8 @@ class ChtWindow(Adw.ApplicationWindow):
         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: