v0.100.14

benchmark/README.md

@@ -26,6 +26,13 @@ without any human review or supervision! The LLM could generate dangerous python
 that harms your system, like this: `import os; os.system("sudo rm -rf /")`.
 Running inside a docker container helps limit the damage that could be done.
 
+The current harness (`benchmark.py`) organizes the exercises as "Cats": each
+exercise carries `cat.yaml` metadata, enabling per-test selection and
+deterministic subsets (e.g. `--hash-re '^0'`) for spot-checking models. The
+canonical exercises source is
+[ErichBSchulz/cecli-cats](https://github.com/ErichBSchulz/cecli-cats), a
+Cats-formatted conversion of the polyglot benchmark.
+
 ## Usage
 
 There are 3 main tasks involved in benchmarking:
@@ -41,17 +48,18 @@ There are 3 main tasks involved in benchmarking:
 These steps only need to be done once.
 
 ```
-ORG=Aider-AI
-REPO=aider
+ORG=cecli-dev
+REPO=cecli
 # Clone the main repo
 git clone https://github.com/$ORG/$REPO.git
 
 # Create the scratch dir to hold benchmarking results inside the main repo:
 cd $REPO
 mkdir tmp.benchmarks
 
-# Clone the repo with the exercises
-git clone https://github.com/$ORG/polyglot-benchmark tmp.benchmarks/polyglot-benchmark
+# Clone the canonical exercises source for the Cats harness
+git clone https://github.com/ErichBSchulz/cecli-cats tmp.benchmarks/cecli-cats
+
 
 # Build the docker container
 ./benchmark/docker_build.sh
@@ -69,10 +77,10 @@ Launch the docker container and run the benchmark inside it:
 ./benchmark/docker.sh
 
 # Run the benchmark:
-./benchmark/benchmark.py a-helpful-name-for-this-run --model gpt-3.5-turbo --edit-format whole --threads 10 --exercises-dir polyglot-benchmark
+./benchmark/benchmark.py a-helpful-name-for-this-run --model gpt-3.5-turbo --edit-format whole --threads 10 --exercises-dir cecli-cats
 
 # Or with OpenRouter models (requires OPENROUTER_API_KEY environment variable):
-./benchmark/benchmark.py openrouter-run --model openrouter/deepseek/deepseek-r1:free --edit-format whole --threads 10 --exercises-dir polyglot-benchmark
+./benchmark/benchmark.py openrouter-run --model openrouter/deepseek/deepseek-r1:free --edit-format whole --threads 10 --exercises-dir cecli-cats
 ```
 
 The above will create a folder
@@ -156,6 +164,7 @@ Note the roadmap priorities:
 
 ## Limitations
 
+- `benchmark_classic.py` is deprecated (it predates the async Coder API).
 - These scripts are not intended for use by typical `cecli` end users.
 - Some of the old (?deprecated) tools are written as `bash` scripts, so it will be hard to use
   them on Windows.

cecli/__init__.py

@@ -1,6 +1,6 @@
 from packaging import version
 
-__version__ = "0.100.11.dev"
+__version__ = "0.100.14.dev"
 safe_version = __version__
 
 try:

cecli/args.py

@@ -598,6 +598,12 @@ def get_parser(default_config_files, git_root):
         default=True,
         help="Enable/disable streaming responses (default: True)",
     )
+    group.add_argument(
+        "--spinner",
+        action=argparse.BooleanOptionalAction,
+        default=True,
+        help="Enable/disable the spinner while waiting for LLM responses (default: True)",
+    )
     group.add_argument(
         "--user-input-color",
         default="#00cc00",

cecli/coders/agent_coder.py

@@ -305,6 +305,25 @@ async def _exec_async():
                 call_result = await litellm.experimental_mcp_client.call_openai_tool(
                     session=session, openai_tool=tool_call_dict
                 )
+            except Exception as e:
+                if server.is_session_expired_error(e):
+                    try:
+                        session = await server.reconnect()
+                        call_result = await litellm.experimental_mcp_client.call_openai_tool(
+                            session=session, openai_tool=tool_call_dict
+                        )
+                    except Exception as retry_exc:
+                        self.io.tool_warning(
+                            f"Executing {tool_name} on {server.name} failed after reconnect:\n"
+                            f"Error: {retry_exc}"
+                        )
+                        return f"Error executing tool call {tool_name}: {retry_exc}"
+                else:
+                    self.io.tool_warning(
+                        f"Executing {tool_name} on {server.name} failed:\nError: {e}"
+                    )
+                    return f"Error executing tool call {tool_name}: {e}"
+            try:
                 content_parts = []
                 if call_result.content:
                     for item in call_result.content:

cecli/coders/base_coder.py

@@ -221,6 +221,7 @@ def total_cached_tokens(self, value):
     message_tokens_sent = 0
     message_tokens_received = 0
     message_cached_tokens = 0
+    message_cost_deferred = None
     add_cache_headers = False
     cache_warming_thread = None
     num_cache_warming_pings = 0
@@ -1497,6 +1498,10 @@ async def _run_linear(self, with_message=None, preproc=True):
                         self.show_announcements()
                     self.suppress_announcements_for_next_prompt = True
 
+                    if self.message_cost_deferred and not self.io.spinner_active:
+                        self.io.tool_output(self.message_cost_deferred)
+                        self.message_cost_deferred = None
+
                     await self.io.recreate_input()
                     await self.io.input_task
                     user_message = self.io.input_task.result()
@@ -1645,6 +1650,10 @@ async def input_task(self, preproc):
                         self.show_announcements()
                     self.suppress_announcements_for_next_prompt = True
 
+                    if self.message_cost_deferred and not self.io.spinner_active:
+                        self.io.tool_output(self.message_cost_deferred)
+                        self.message_cost_deferred = None
+
                     # Stop spinner before showing announcements or getting input
                     self.io.stop_spinner()
                     self.copy_context()
@@ -2512,7 +2521,11 @@ async def format_in_executor():
             if not self.tui:
                 spinner_text += f" • ${self.format_cost(self.total_cost)} session"
 
-            self.io.start_spinner(spinner_text, coder_uuid=getattr(self, "uuid", None))
+            if self.io.spinner_active:
+                self.io.start_spinner(spinner_text, coder_uuid=getattr(self, "uuid", None))
+            else:
+                self.message_cost_deferred = spinner_text
+
             if self.stream:
                 self.mdstream = True
             else:
@@ -2635,6 +2648,10 @@ async def format_in_executor():
 
             # Ensure any waiting spinner is stopped
             self.io.start_spinner("Processing Answer...", coder_uuid=getattr(self, "uuid", None))
+
+            if not self.io.spinner_active:
+                self.partial_response_content = self.get_multi_response_content_in_progress(True)
+
             self.remove_reasoning_content()
             self.multi_response_content = ""
 
@@ -2980,12 +2997,22 @@ async def _execute_mcp_tools(self, server, tool_calls):
                             continue
 
                         async def do_tool_call():
+                            nonlocal session
                             from litellm import experimental_mcp_client
 
-                            return await experimental_mcp_client.call_openai_tool(
-                                session=session,
-                                openai_tool=new_tool_call,
-                            )
+                            try:
+                                return await experimental_mcp_client.call_openai_tool(
+                                    session=session,
+                                    openai_tool=new_tool_call,
+                                )
+                            except Exception as e:
+                                if server.is_session_expired_error(e):
+                                    session = await server.reconnect()
+                                    return await experimental_mcp_client.call_openai_tool(
+                                        session=session,
+                                        openai_tool=new_tool_call,
+                                    )
+                                raise
 
                         call_result, interrupted = await coroutines.interruptible(
                             do_tool_call(), self.interrupt_event

cecli/interruptible_input.py

@@ -17,7 +17,14 @@ def __init__(self):
             raise RuntimeError("InterruptibleInput is Unix-only (requires selectable stdin).")
 
         self._cancel = threading.Event()
-        self._sel = selectors.DefaultSelector()
+
+        # The default selector (Kqueue on macOS, Epoll on Linux) cannot
+        # handle pipe-based stdin (e.g. when running inside Emacs comint-mode).
+        # Fall back to SelectSelector which works with any fd that supports select().
+        if not sys.stdin.isatty():
+            self._sel = selectors.SelectSelector()
+        else:
+            self._sel = selectors.DefaultSelector()
 
         # self-pipe to wake up select() from interrupt()
         self._r, self._w = os.pipe()

cecli/io.py

@@ -369,6 +369,7 @@ def __init__(
         notifications_command=None,
         notification_bell=False,
         verbose=False,
+        show_spinner=True,
     ):
         self.console = Console()
         self.pretty = pretty
@@ -499,6 +500,7 @@ def __init__(
             fancy_input = False
 
         # Spinner state
+        self.spinner_active = show_spinner
         self.spinner_running = False
         self.spinner_text = ""
         self.last_spinner_text = ""
@@ -507,7 +509,7 @@ def __init__(
         self.spinner_last_frame_index = 0
         self.unicode_palette = "░█"
         self.fallback_spinner = None
-        self.fallback_spinner_enabled = True
+        self.fallback_spinner_enabled = show_spinner
 
         self.interruptible_input = None
 
@@ -569,7 +571,12 @@ def start_spinner(self, text, update_last_text=True, **kwargs):
         """Start the spinner."""
         self.stop_spinner()
 
+        if not self.spinner_active:
+            return
+
         if self.prompt_session:
+            if not self.fallback_spinner_enabled:
+                return
             self.spinner_running = True
             self.spinner_text = text
             self.spinner_frame_index = self.spinner_last_frame_index
@@ -582,9 +589,15 @@ def start_spinner(self, text, update_last_text=True, **kwargs):
             self.fallback_spinner.step()
 
     def update_spinner(self, text):
+        if not self.spinner_active:
+            return
+
         self.spinner_text = text
 
     def update_spinner_suffix(self, text=None):
+        if not self.spinner_active:
+            return
+
         if text:
             self.spinner_suffix = f" • {text[:16].strip()}"
         else:

cecli/main.py

@@ -40,6 +40,20 @@
 if sys.platform == "win32":
     if hasattr(asyncio, "set_event_loop_policy"):
         asyncio.set_event_loop_policy(asyncio.WindowsSelectorEventLoopPolicy())
+elif sys.platform == "darwin":
+    # The default KqueueSelector cannot handle pipe-based stdin
+    # (e.g. when running inside Emacs comint-mode). Fall back to
+    # SelectSelector which works with any file descriptor that supports select().
+    import selectors
+
+    if not sys.stdin.isatty():
+        _original_event_loop_policy = asyncio.DefaultEventLoopPolicy
+
+        class _SelectSelectorPolicy(asyncio.DefaultEventLoopPolicy):
+            def new_event_loop(self):
+                return asyncio.SelectorEventLoop(selectors.SelectSelector())
+
+        asyncio.set_event_loop_policy(_SelectSelectorPolicy())
 from prompt_toolkit.enums import EditingMode
 
 from .dump import dump  # noqa
@@ -708,6 +722,7 @@ def get_io(pretty):
             notifications_command=args.notifications_command,
             notification_bell=args.notification_bell,
             verbose=args.verbose,
+            show_spinner=args.spinner,
         )
 
     validate_tui_args(args)

cecli/mcp/manager.py

@@ -160,17 +160,46 @@ async def connect_server(self, name: str) -> bool:
             self._server_tools[server.name] = get_local_tool_schemas()
             return True
 
-        try:
-            session = await server.connect()
-            tools = await experimental_mcp_client.load_mcp_tools(session=session, format="openai")
-            self._server_tools[server.name] = tools
-            self._connected_servers.add(server)
-            self._log_verbose(f"Connected to MCP server: {name}")
-            return True
-        except (Exception, asyncio.CancelledError) as e:
-            if server.name != "unnamed-server":
-                self._log_error(f"Failed to connect to MCP server {name}: {e}")
-            return False
+        # Retry with exponential backoff for transient connection failures.
+        # Note: This also fixes a latent bug where asyncio.CancelledError was
+        # silently caught and treated as a connection failure. CancelledError is
+        # now re-raised to properly propagate cancellation.
+        # When io is None (e.g., during from_servers before IO is assigned),
+        # _log_warning and _log_error silently return — retries still happen
+        # but with no user-visible feedback. This is intentional.
+        max_retries = 3 if server.name != "unnamed-server" else 1
+        delay = 1.0
+        backoff = 2.0
+        max_delay = 30.0
+
+        for attempt in range(1, max_retries + 1):
+            try:
+                session = await server.connect()
+                tools = await experimental_mcp_client.load_mcp_tools(
+                    session=session, format="openai"
+                )
+                self._server_tools[server.name] = tools
+                self._connected_servers.add(server)
+                self._log_verbose(f"Connected to MCP server: {name}")
+                return True
+            except asyncio.CancelledError:
+                raise
+            except Exception as e:
+                if attempt < max_retries and server.name != "unnamed-server":
+                    self._log_warning(
+                        f"Connection attempt {attempt} failed for {name}, "
+                        f"retrying in {delay}s... ({e})"
+                    )
+
+                    await asyncio.sleep(delay)
+                    delay = min(delay * backoff, max_delay)
+                else:
+                    if server.name != "unnamed-server":
+                        self._log_error(
+                            f"Failed to connect to MCP server {name} "
+                            f"after {max_retries} attempts: {e}"
+                        )
+                    return False
 
     async def disconnect_server(self, name: str) -> bool:
         """
@@ -281,11 +310,10 @@ async def add_server_with_retry(
                 success = await mcp_manager.add_server(server, connect=False)
                 return (server, success)
 
-            for _attempt in range(max_retries):
-                success = await mcp_manager.add_server(server, connect=True)
-                if success:
-                    return (server, True)
-            return (server, False)
+            # connect_server now has built-in retry logic, so we only need
+            # a single call here — no separate retry loop needed.
+            success = await mcp_manager.add_server(server, connect=True)
+            return (server, success)
 
         tasks = []
         for server in servers: