V1.1.0

cecli/coders/agent_coder.py

@@ -27,7 +27,6 @@
 )
 from cecli.helpers.skills import SkillsManager
 from cecli.hooks import HookIntegration
-from cecli.llm import litellm
 from cecli.mcp import LocalServer, McpServerManager
 from cecli.tools.utils.base_tool import BaseTool
 from cecli.tools.utils.registry import ToolRegistry
@@ -364,16 +363,12 @@ async def _exec_async():
             }
             try:
                 session = await server.connect()
-                call_result = await litellm.experimental_mcp_client.call_openai_tool(
-                    session=session, openai_tool=tool_call_dict
-                )
+                call_result = await self.call_mcp_tool_from_session(session, 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
-                        )
+                        call_result = await self.call_mcp_tool_from_session(session, tool_call_dict)
                     except Exception as retry_exc:
                         self.io.tool_warning(
                             f"Executing {tool_name} on {server.name} failed after reconnect:\n"

cecli/commands/add.py

@@ -132,7 +132,24 @@ async def execute(cls, io, coder, args, **kwargs):
                         f" {active_model.name} does not support images."
                     )
                     continue
-                content = io.read_text(abs_file_path)
+                try:
+                    content = io.read_text(abs_file_path)
+                except (ValueError, UnicodeError, UnicodeDecodeError, OSError) as exc:
+                    # Binary or undecodable files (e.g. .git/objects/pack/*.rev)
+                    # raise ValueError("Could not determine text encoding ...")
+                    # from decoding.safe_open. Skip them gracefully instead of
+                    # crashing the session.
+                    msg = str(exc)
+                    if "Could not determine text encoding" in msg or isinstance(
+                        exc, (UnicodeError, UnicodeDecodeError)
+                    ):
+                        io.tool_warning(
+                            f"Skipping {matched_file}: not decodable as text "
+                            "(binary or unknown encoding)"
+                        )
+                    else:
+                        io.tool_error(f"Skipping {matched_file}: {exc}")
+                    continue
                 if content is None:
                     io.tool_error(f"Unable to read {matched_file}")
                 else:

cecli/commands/reasoning_effort.py

@@ -14,7 +14,7 @@ class ReasoningEffortCommand(BaseCommand):
     @classmethod
     async def execute(cls, io, coder, args, **kwargs):
         """Execute the reasoning-effort command with given parameters."""
-        model = coder.main_model
+        model = coder.get_active_model()
 
         if not args.strip():
             # Display current value if no args are provided

cecli/commands/think_tokens.py

@@ -11,7 +11,7 @@ class ThinkTokensCommand(BaseCommand):
     @classmethod
     async def execute(cls, io, coder, args, **kwargs):
         """Execute the think-tokens command with given parameters."""
-        model = coder.main_model
+        model = coder.get_active_model()
 
         if not args.strip():
             # Display current value if no args are provided

cecli/helpers/io_proxy.py

@@ -51,6 +51,9 @@ def __init__(self, target: T, coder: Any) -> None:
         super().__setattr__("_coder", weakref.ref(coder))
         # Per-coder task storage: {coder_uuid: {attr_name: asyncio.Task}}
         super().__setattr__("_per_coder", {coder_uuid: {}})
+        # Last tool `type` emitted via tool_output — lives on the proxy,
+        # never on the shared target (like coder_uuid)
+        super().__setattr__("_last_type", None)
 
         # Register a per-coder input queue (TUI mode only)
         # Allows the TUI to push input directly to this coder's queue,
@@ -76,6 +79,7 @@ def tool_output(self, *messages: Any, **kwargs: Any) -> Any:
         """Forward tool_output with coder_uuid injected."""
         if "coder_uuid" not in kwargs:
             kwargs["coder_uuid"] = self._coder_uuid
+        self._last_type = kwargs.get("type")
         return self._target.tool_output(*messages, **kwargs)
 
     def tool_error(self, message: str = "", strip: bool = True, **kwargs: Any) -> Any:
@@ -265,7 +269,7 @@ def __getattr__(self, name: str) -> Any:
 
     def __setattr__(self, name: str, value: Any) -> None:
         # Proxy-internal attributes — store on proxy instance only
-        if name in ("_target", "_coder_uuid", "_coder", "_per_coder"):
+        if name in ("_target", "_coder_uuid", "_coder", "_per_coder", "_last_type"):
             super().__setattr__(name, value)
         # Per-coder task attributes — isolate per-coder so coders don't
         # compete for the same promise on the shared InputOutput instance

cecli/helpers/model_config/__init__.py

@@ -0,0 +1,10 @@
+"""Derive default per-model configuration from model metadata.
+
+The model config package turns a flat litellm-style model metadata file into the
+``{api, llm, agent}`` override blocks that :mod:`cecli.models` consumes,
+mirroring the ``model-overrides`` section of ``.cecli.conf.yml``.
+"""
+
+from .pipeline import ModelConfigPipeline, get_default_config
+
+__all__ = ["ModelConfigPipeline", "get_default_config"]

cecli/helpers/model_config/agent.py

@@ -0,0 +1,40 @@
+"""Derive the ``agent`` override block for a model.
+
+The agent block holds ModelSettings overrides. :mod:`cecli.models` applies each
+of these directly (``setattr``) the same way the ``agent`` section of a
+``model-overrides`` entry is applied.
+"""
+
+from __future__ import annotations
+
+from typing import Dict, Optional
+
+from .identifiers import is_anthropic
+from .utils import supports_reasoning
+
+
+def derive_agent_config(provider: Optional[str], route: str, record: Optional[Dict]) -> Dict:
+    """Return the ``agent`` config block for a model.
+
+    Args:
+        provider: Provider portion of the model name (may be ``None``).
+        route: Model route (name after the provider prefix).
+        record: The matched model metadata record, or ``None`` for unknown models.
+
+    Returns:
+        A dict of ModelSettings overrides (caching, temperature handling).
+    """
+    reasoning = supports_reasoning(record)
+    record = record or {}
+    agent: Dict = {
+        "cache_control": is_anthropic(provider, route, record),
+        # ``cache_read_input_token_cost`` in the metadata is the determinant for
+        # whether a model supports prompt caching. Unknown models default to
+        # assuming caching support.
+        "caches_by_default": bool(record.get("cache_read_input_token_cost")) if record else True,
+    }
+
+    if reasoning or record.get("supports_adaptive_thinking"):
+        agent["use_temperature"] = False
+
+    return agent

cecli/helpers/model_config/api.py

@@ -0,0 +1,62 @@
+"""Derive the ``api`` override block for a model.
+
+The api block holds request-level parameters. :mod:`cecli.models` merges each
+of these keys into ``extra_params`` the same way the ``api`` section of a
+``model-overrides`` entry is applied.
+"""
+
+from __future__ import annotations
+
+from typing import Dict, Optional
+
+from .identifiers import is_anthropic, is_claude_5_plus, is_gemini_2_5
+from .utils import supports_reasoning
+
+_THINKING_BUDGET_TOKENS = 2048
+#: Default thinking budget for Gemini 2.5 models (Gemini 2.5 Pro's default).
+_GEMINI_THINKING_BUDGET_TOKENS = 8192
+
+
+def derive_api_config(provider: Optional[str], route: str, record: Optional[Dict]) -> Dict:
+    """Return the ``api`` config block for a model.
+
+    Args:
+        provider: Provider portion of the model name (may be ``None``).
+        route: Model route (name after the provider prefix).
+        record: The matched model metadata record, or ``None`` for unknown models.
+
+    Returns:
+        A dict of request-level params (reasoning format, thinking, tool calls).
+    """
+    reasoning = supports_reasoning(record)
+    record = record or {}
+    gemini_2_5 = is_gemini_2_5(provider, route, record)
+    api: Dict = {}
+
+    if reasoning and not gemini_2_5:
+        effort = _default_reasoning_effort(record)
+
+        if effort:
+            api["reasoning_effort"] = effort
+
+    if is_anthropic(provider, route, record) and not is_claude_5_plus(provider, route, record):
+        # Claude 5+ uses adaptive thinking via ``reasoning_effort`` instead of
+        # the ``thinking.type.enabled`` budget block.
+        api["thinking"] = {"type": "enabled", "budget_tokens": _THINKING_BUDGET_TOKENS}
+    elif gemini_2_5:
+        # Gemini 2.5 configures thinking via a token budget; litellm maps the
+        # generic ``thinking`` param onto ``thinkingBudget`` + ``includeThoughts``.
+        api["thinking"] = {"type": "enabled", "budget_tokens": _GEMINI_THINKING_BUDGET_TOKENS}
+
+    if record.get("supports_parallel_function_calling", True):
+        api["parallel_tool_calls"] = True
+
+    return api
+
+
+def _default_reasoning_effort(record):
+    """Default reasoning effort for a reasoning-capable model.
+
+    Always ``medium``; the metadata effort flags are intentionally not used.
+    """
+    return "medium"

cecli/helpers/model_config/formatters/__init__.py

@@ -0,0 +1,20 @@
+"""Provider-specific helper overrides for the model config pipeline."""
+
+from .reasoning import anthropic_reasoning, format_reasoning, gemini_reasoning, noop
+from .thinking import (
+    anthropic_5_thinking,
+    anthropic_thinking,
+    format_thinking,
+    gemini_thinking,
+)
+
+__all__ = [
+    "format_reasoning",
+    "anthropic_reasoning",
+    "gemini_reasoning",
+    "noop",
+    "format_thinking",
+    "anthropic_thinking",
+    "anthropic_5_thinking",
+    "gemini_thinking",
+]

cecli/helpers/model_config/formatters/reasoning.py

@@ -0,0 +1,83 @@
+"""Reasoning formatters for the model config pipeline.
+
+These helpers rewrite the generic reasoning effort shape onto the params
+litellm understands for a provider.  ``extra_params`` are the kwargs passed to
+``litellm.acompletion``, so a formatter exposes the effort as a top-level
+litellm param and lets litellm map it onto the provider's own field.  For
+example, Gemini uses ``thinkingConfig`` under the hood: litellm maps a
+top-level ``reasoning_effort`` to ``thinkingLevel`` (Gemini 3) or
+``thinkingBudget`` (Gemini 2.5).  ``set_reasoning_effort`` in
+:mod:`cecli.models` invokes the formatter chosen by the pipeline
+(``helpers.format_reasoning``) after it has applied the generic shape.
+"""
+
+from __future__ import annotations
+
+from typing import Callable, Dict, Optional
+
+from ..identifiers import is_claude_5_plus, is_gemini
+
+
+def format_reasoning(provider: Optional[str], route: str, record: Optional[Dict]) -> Callable:
+    """Return the reasoning formatter for a model.
+
+    Args:
+        provider: Provider portion of the model name (may be ``None``).
+        route: Model route (name after the provider prefix).
+        record: The matched model metadata record, or ``None`` for unknown models.
+
+    Returns:
+        A callable that mutates ``extra_params`` in place.  Unknown models get
+        a noop so ``set_reasoning_effort`` keeps its default behavior.
+    """
+    if is_gemini(provider, route, record):
+        return gemini_reasoning
+
+    if is_claude_5_plus(provider, route, record):
+        return anthropic_reasoning
+
+    return noop
+
+
+def noop(extra_params: Dict) -> Dict:
+    """Default formatter: leave ``extra_params`` untouched."""
+    return extra_params
+
+
+def gemini_reasoning(extra_params: Dict) -> Dict:
+    """Gemini models configure thinking via litellm's ``reasoning_effort``.
+
+    litellm maps the top-level ``reasoning_effort`` kwarg onto Gemini's
+    ``thinkingConfig`` (``thinkingLevel`` for Gemini 3, ``thinkingBudget`` for
+    Gemini 2.5) and sets ``includeThoughts``, so the generic effort is lifted
+    out of ``extra_body``.
+    """
+    return _lift_reasoning_effort(extra_params)
+
+
+def anthropic_reasoning(extra_params: Dict) -> Dict:
+    """Claude 5+ models configure thinking via litellm's ``reasoning_effort``.
+
+    litellm maps the top-level ``reasoning_effort`` kwarg onto
+    ``thinking.type.adaptive`` + ``output_config.effort``, so the generic
+    effort is lifted out of ``extra_body`` and ``extra_body`` is dropped
+    (Anthropic does not accept extra inputs).
+    """
+    params = _lift_reasoning_effort(extra_params)
+    params.pop("extra_body", None)
+    return params
+
+
+def _lift_reasoning_effort(extra_params: Dict) -> Dict:
+    """Move the generic ``reasoning_effort`` out of ``extra_body`` to top level."""
+    extra_body = extra_params.get("extra_body")
+
+    if not isinstance(extra_body, dict):
+        return extra_params
+
+    effort = extra_body.pop("reasoning_effort", None)
+
+    if effort is not None:
+        extra_params["reasoning_effort"] = effort
+
+    return extra_params