agent improvement

cht/agent/claude_sdk_connection.py

@@ -37,9 +37,11 @@ You help the user understand what happened during their recording session.
 You have access to frame screenshots extracted from the recording. When frames are mentioned,
 use the Read tool to view them. Frame timestamps are in seconds from the start of the recording.
 
-You also have tools to search transcripts, get session info, and capture new frames.
+You can use any available tools including WebFetch and WebSearch when the user asks you to
+look something up. Use them freely — all tools are pre-authorized.
 
-Be concise and specific. Focus on what's visible in the frames."""
+Your primary role is description and analysis, not code generation. Be concise and specific.
+Focus on what's visible in the frames and what's in the transcript."""
 
 MODELS = [
     "claude-sonnet-4-6",
@@ -78,21 +80,16 @@ def _messages_to_prompt(messages: list[Message]) -> str:
     return "\n".join(lines)
 
 
-def _tool_schemas(tools: list[Tool]) -> list[str]:
-    """Extract tool names for the SDK's allowed_tools parameter."""
-    # The Claude SDK uses allowed_tools as a list of tool name strings.
-    # Our custom tools are executed by the runner, not by the SDK,
-    # so we only pass "Read" to the SDK (for frame viewing).
-    return ["Read"]
-
-
 class ClaudeSDKConnection:
     """AgentConnection using claude_agent_sdk — requires Claude Code CLI."""
 
-    def __init__(self, cwd: str | None = None, max_turns: int = 5, model: str = MODELS[0]):
+    def __init__(self, cwd: str | None = None, max_turns: int | None = None,
+                 model: str = MODELS[0], permission_mode: str | None = None):
+        from cht.config import AGENT_PERMISSION_MODE, AGENT_MAX_TURNS
         self._cwd = cwd
-        self._max_turns = max_turns
+        self._max_turns = max_turns or AGENT_MAX_TURNS
         self._model = model
+        self._permission_mode = permission_mode or AGENT_PERMISSION_MODE
         self._cancelled = False
 
     @property
@@ -147,9 +144,9 @@ class ClaudeSDKConnection:
                     options=ClaudeAgentOptions(
                         model=self._model,
                         cwd=cwd or ".",
-                        allowed_tools=_tool_schemas(tools),
                         system_prompt=SYSTEM_PROMPT,
                         max_turns=self._max_turns,
+                        permission_mode=self._permission_mode,
                     ),
                 ):
                     if self._cancelled:

cht/agent/runner.py

@@ -177,6 +177,20 @@ class AgentRunner:
     def model(self, value: str):
         self._get_connection().set_model(value)
 
+    @property
+    def permission_mode(self) -> str:
+        conn = self._get_connection()
+        return getattr(conn, "_permission_mode", "default")
+
+    @permission_mode.setter
+    def permission_mode(self, value: str):
+        conn = self._get_connection()
+        if hasattr(conn, "_permission_mode"):
+            conn._permission_mode = value
+            import cht.config
+            cht.config.AGENT_PERMISSION_MODE = value
+            log.info("Permission mode set to %s", value)
+
     def clear_history(self):
         self._thread = AgentThread()
 

cht/ui/agent_input.py

@@ -21,6 +21,7 @@ class AgentInputPanel(Gtk.Frame):
         "model-changed": (GObject.SignalFlags.RUN_FIRST, None, (str,)),
         "lang-changed": (GObject.SignalFlags.RUN_FIRST, None, (str,)),
         "history-toggled": (GObject.SignalFlags.RUN_FIRST, None, (bool,)),
+        "permission-changed": (GObject.SignalFlags.RUN_FIRST, None, (str,)),
     }
 
     def __init__(self, **kwargs):
@@ -68,6 +69,18 @@ class AgentInputPanel(Gtk.Frame):
         history_toggle.connect("toggled", lambda b: self.emit("history-toggled", b.get_active()))
         actions_box.append(history_toggle)
 
+        perm_label = Gtk.Label(label="Perms:")
+        perm_label.add_css_class("dim-label")
+        actions_box.append(perm_label)
+
+        self._perm_modes = ["default", "acceptEdits", "bypassPermissions", "dontAsk", "plan"]
+        self._perm_labels = ["Default", "Accept Edits", "Bypass All", "Don't Ask", "Plan Only"]
+        self._perm_dropdown = Gtk.DropDown.new_from_strings(self._perm_labels)
+        self._perm_dropdown.set_size_request(140, -1)
+        self._perm_dropdown.set_tooltip_text("Claude SDK permission mode")
+        self._perm_dropdown.connect("notify::selected", self._on_perm_changed)
+        actions_box.append(self._perm_dropdown)
+
         outer.append(actions_box)
 
         input_row = Gtk.Box(orientation=Gtk.Orientation.HORIZONTAL, spacing=4)
@@ -123,3 +136,15 @@ class AgentInputPanel(Gtk.Frame):
         if idx < len(self._lang_names):
             lang_code = LANGUAGES[self._lang_names[idx]]
             self.emit("lang-changed", lang_code or "")
+
+    def _on_perm_changed(self, dropdown, _pspec):
+        idx = dropdown.get_selected()
+        if idx < len(self._perm_modes):
+            self.emit("permission-changed", self._perm_modes[idx])
+
+    def set_permission_mode(self, mode: str) -> None:
+        try:
+            idx = self._perm_modes.index(mode)
+            self._perm_dropdown.set_selected(idx)
+        except ValueError:
+            pass

cht/ui/agent_output.py

@@ -1,4 +1,4 @@
-"""Agent output panel — single TextView, fully selectable/copy-pastable."""
+"""Agent output panel — single TextView, colored message regions, copy-pastable."""
 
 import logging
 
@@ -41,6 +41,11 @@ class AgentOutputPanel(Gtk.Frame):
         label.set_halign(Gtk.Align.START)
         header.append(label)
 
+        # Spinner for "thinking" feedback
+        self._spinner = Gtk.Spinner()
+        self._spinner.set_visible(False)
+        header.append(self._spinner)
+
         clear_btn = Gtk.Button(label="Clear")
         clear_btn.add_css_class("flat")
         clear_btn.connect("clicked", lambda _: self.clear())
@@ -66,52 +71,72 @@ class AgentOutputPanel(Gtk.Frame):
 
         self.set_child(box)
 
-        # Streaming state: track where the current assistant response starts
+        # Streaming state
         self._response_marks: dict[str, Gtk.TextMark] = {}
         self._response_accum: dict[str, list[str]] = {}
 
     def _setup_tags(self):
         buf = self._view.get_buffer()
         markdown.setup_tags(buf)
+        # Message region backgrounds — user indented right, assistant left
+        buf.create_tag("user-bg",
+                       paragraph_background="#1a2a3a",
+                       left_margin=120, right_margin=12,
+                       pixels_above_lines=6, pixels_below_lines=6)
+        buf.create_tag("assistant-bg",
+                       paragraph_background="#1a2a1a",
+                       left_margin=12, right_margin=120,
+                       pixels_above_lines=4, pixels_below_lines=4)
+        # Inline styles
         buf.create_tag("user-prefix", weight=700, foreground="#7aafff")
+        buf.create_tag("user-text", foreground="#c8ddf0")
         buf.create_tag("assistant-prefix", weight=700, foreground="#8bc78b")
         buf.create_tag("tool-prefix", weight=700, foreground="#d4a053")
-        buf.create_tag("tool-output", foreground="#aaaaaa", left_margin=16)
+        buf.create_tag("tool-output", foreground="#aaaaaa", left_margin=24)
         buf.create_tag("error", foreground="#ff6b6b")
-        buf.create_tag("ref-chip", foreground="#7aafff", style=Pango.Style.ITALIC)
+        buf.create_tag("ref-chip", foreground="#5a9fd4", style=Pango.Style.ITALIC)
         buf.create_tag("status", foreground="#888888")
+        buf.create_tag("thinking", foreground="#666666", style=Pango.Style.ITALIC)
+
+    # -- helpers --
+
+    def _insert_tagged(self, text: str, *tags: str):
+        """Insert text at end and apply tags."""
+        buf = self._view.get_buffer()
+        mark = buf.create_mark(None, buf.get_end_iter(), True)
+        buf.insert(buf.get_end_iter(), text)
+        start = buf.get_iter_at_mark(mark)
+        end = buf.get_end_iter()
+        for tag in tags:
+            buf.apply_tag_by_name(tag, start, end)
+        buf.delete_mark(mark)
 
     # -- Public API --
 
     def append(self, text: str) -> None:
         """Append a status/info line."""
-        buf = self._view.get_buffer()
+        self._insert_tagged(text, "status")
-        end = buf.get_end_iter()
-        mark = buf.create_mark(None, end, True)
-        buf.insert(end, text)
-        start = buf.get_iter_at_mark(mark)
-        buf.apply_tag_by_name("status", start, buf.get_end_iter())
-        buf.delete_mark(mark)
         self._scroll_to_bottom()
 
     def clear(self) -> None:
         self._view.get_buffer().set_text("")
         self._response_marks.clear()
         self._response_accum.clear()
+        self._spinner.set_visible(False)
+        self._spinner.stop()
 
     def add_user_message(self, text: str, frames: list | None = None,
                          transcripts: list | None = None) -> None:
         buf = self._view.get_buffer()
-        end = buf.get_end_iter()
+
+        # Region start mark
+        region_mark = buf.create_mark(None, buf.get_end_iter(), True)
 
         # Prefix
-        mark = buf.create_mark(None, end, True)
+        self._insert_tagged("\n> ", "user-prefix")
-        buf.insert(end, "\n> ")
-        buf.apply_tag_by_name("user-prefix", buf.get_iter_at_mark(mark), buf.get_end_iter())
-        buf.delete_mark(mark)
 
         # Text
-        buf.insert(buf.get_end_iter(), text)
+        self._insert_tagged(text, "user-text")
 
         # Reference chips
         refs = []
@@ -124,13 +149,15 @@ class AgentOutputPanel(Gtk.Frame):
                 tid = t.transcript_id if hasattr(t, "transcript_id") else (t.id if hasattr(t, "id") else str(t))
                 refs.append(tid)
         if refs:
-            end = buf.get_end_iter()
+            self._insert_tagged(" [" + ", ".join(refs) + "]", "ref-chip")
-            mark = buf.create_mark(None, end, True)
-            buf.insert(end, " [" + ", ".join(refs) + "]")
-            buf.apply_tag_by_name("ref-chip", buf.get_iter_at_mark(mark), buf.get_end_iter())
-            buf.delete_mark(mark)
 
         buf.insert(buf.get_end_iter(), "\n")
+
+        # Apply background to entire user region
+        buf.apply_tag_by_name("user-bg",
+                              buf.get_iter_at_mark(region_mark),
+                              buf.get_end_iter())
+        buf.delete_mark(region_mark)
         self._scroll_to_bottom()
 
     def begin_assistant_message(self, msg_id: str) -> None:
@@ -141,63 +168,73 @@ class AgentOutputPanel(Gtk.Frame):
         self._response_marks[msg_id] = buf.create_mark(f"resp_{msg_id}", end, True)
         self._response_accum[msg_id] = []
 
+        # Show inline thinking indicator + header spinner
+        self._insert_tagged("thinking...\n", "thinking")
+        self._spinner.set_visible(True)
+        self._spinner.start()
+        self._scroll_to_bottom()
+
     def append_to_assistant(self, msg_id: str, text: str) -> None:
         if msg_id not in self._response_accum:
             return
-        self._response_accum[msg_id].append(text)
         buf = self._view.get_buffer()
+        # On first chunk, clear the "thinking..." placeholder
+        if not self._response_accum[msg_id]:
+            mark = self._response_marks.get(msg_id)
+            if mark:
+                start = buf.get_iter_at_mark(mark)
+                buf.delete(start, buf.get_end_iter())
+        self._response_accum[msg_id].append(text)
         buf.insert(buf.get_end_iter(), text)
         self._scroll_to_bottom()
 
     def finish_assistant(self, msg_id: str, full_text: str) -> None:
+        # Stop spinner
+        self._spinner.set_visible(False)
+        self._spinner.stop()
+
         mark = self._response_marks.pop(msg_id, None)
         self._response_accum.pop(msg_id, None)
         if not mark:
             return
+
         buf = self._view.get_buffer()
         start = buf.get_iter_at_mark(mark)
         end = buf.get_end_iter()
         buf.delete(start, end)
+
+        # Re-render with markdown
+        region_start = buf.create_mark(None, buf.get_iter_at_mark(mark), True)
         it = buf.get_iter_at_mark(mark)
         markdown.render(buf, it, full_text)
         buf.insert(buf.get_end_iter(), "\n")
+
+        # Apply assistant background to rendered region
+        buf.apply_tag_by_name("assistant-bg",
+                              buf.get_iter_at_mark(region_start),
+                              buf.get_end_iter())
+        buf.delete_mark(region_start)
         buf.delete_mark(mark)
 
     def add_tool_call(self, tool_use: ToolUse) -> None:
-        buf = self._view.get_buffer()
+        self._insert_tagged(f"  ▶ {tool_use.tool_name}", "tool-prefix")
-        end = buf.get_end_iter()
-
-        mark = buf.create_mark(None, end, True)
-        buf.insert(end, f"  ▶ {tool_use.tool_name}")
-        buf.apply_tag_by_name("tool-prefix", buf.get_iter_at_mark(mark), buf.get_end_iter())
-        buf.delete_mark(mark)
 
         if tool_use.input:
             inp = str(tool_use.input)
             if len(inp) > 80:
                 inp = inp[:77] + "..."
-            end = buf.get_end_iter()
+            self._insert_tagged(f" {inp}", "tool-output")
-            mark = buf.create_mark(None, end, True)
-            buf.insert(end, f" {inp}")
-            buf.apply_tag_by_name("tool-output", buf.get_iter_at_mark(mark), buf.get_end_iter())
-            buf.delete_mark(mark)
 
-        buf.insert(buf.get_end_iter(), "\n")
+        self._view.get_buffer().insert(self._view.get_buffer().get_end_iter(), "\n")
         self._scroll_to_bottom()
 
     def update_tool_result(self, tool_use_id: str, result: ToolResult) -> None:
-        buf = self._view.get_buffer()
         text = result.error or result.output or ""
         if not text:
             return
-        end = buf.get_end_iter()
-        mark = buf.create_mark(None, end, True)
-        # Indent tool output
         indented = "\n".join(f"    {line}" for line in text.split("\n"))
         tag = "error" if result.error else "tool-output"
-        buf.insert(end, indented + "\n")
+        self._insert_tagged(indented + "\n", tag)
-        buf.apply_tag_by_name(tag, buf.get_iter_at_mark(mark), buf.get_end_iter())
-        buf.delete_mark(mark)
         self._scroll_to_bottom()
 
     def load_thread(self, thread: Thread) -> None:
@@ -219,9 +256,14 @@ class AgentOutputPanel(Gtk.Frame):
             elif isinstance(msg, AssistantMessage):
                 text = " ".join(b.text for b in msg.content if isinstance(b, TextBlock))
                 buf = self._view.get_buffer()
+                region_mark = buf.create_mark(None, buf.get_end_iter(), True)
                 it = buf.get_end_iter()
                 markdown.render(buf, it, text)
                 buf.insert(buf.get_end_iter(), "\n")
+                buf.apply_tag_by_name("assistant-bg",
+                                      buf.get_iter_at_mark(region_mark),
+                                      buf.get_end_iter())
+                buf.delete_mark(region_mark)
             elif isinstance(msg, ToolUse):
                 self.add_tool_call(msg)
             elif isinstance(msg, ToolResult):

cht/window.py

@@ -555,6 +555,7 @@ class ChtWindow(Adw.ApplicationWindow):
         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))
+        self._agent_input.connect("permission-changed", self._on_permission_changed)
         right_box.append(self._agent_input)
 
         return right_box
@@ -678,10 +679,16 @@ class ChtWindow(Adw.ApplicationWindow):
         self._agent.model = model
         log.info("Model switched to %s", model)
 
+    def _on_permission_changed(self, _panel, mode):
+        self._agent.permission_mode = mode
+        log.info("Permission mode switched to %s", mode)
+
     def _populate_model_dropdown(self):
         self._agent_input.populate_models(
             self._agent.available_models, self._agent.model
         )
+        from cht.config import AGENT_PERMISSION_MODE
+        self._agent_input.set_permission_mode(AGENT_PERMISSION_MODE)
 
     def _check_agent_auth(self):
         import os