cecli-dev 1.0.5__tar.gz → 1.1.0rc2__tar.gz

{cecli_dev-1.0.5 → cecli_dev-1.1.0rc2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cecli-dev
-Version: 1.0.5
+Version: 1.1.0rc2
 Summary: we also can't pronounce cecli
 Project-URL: Homepage, https://github.com/dwash96/cecli
 Project-URL: Documentation, https://github.com/dwash96/cecli

{cecli_dev-1.0.5 → cecli_dev-1.1.0rc2}/cecli/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '1.0.5'
-__version_tuple__ = version_tuple = (1, 0, 5)
+__version__ = version = '1.1.0rc2'
+__version_tuple__ = version_tuple = (1, 1, 0, 'rc2')
 
-__commit_id__ = commit_id = 'g7fb4509c5'
+__commit_id__ = commit_id = 'g08300992a'

{cecli_dev-1.0.5 → cecli_dev-1.1.0rc2}/cecli/coders/base_coder.py

@@ -3440,11 +3440,17 @@ class Coder(metaclass=UsageMeta):
         to be `None` when `tool_calls` are present.
         """
         msg = dict(role="assistant")
-        response = (
-            self.partial_response_chunks[0]
-            if not self.stream
-            else litellm.stream_chunk_builder(self.partial_response_chunks)
-        )
+
+        # Prefer the response already produced by consolidate_chunks(): it carries
+        # the provider-specific fields (e.g. reasoning_items) that we preserved
+        # across all chunks, which a fresh litellm.stream_chunk_builder() pass
+        # alone would drop or truncate.
+        if self.partial_response_consolidated:
+            response = self.partial_response_consolidated[0]
+        elif not self.stream:
+            response = self.partial_response_chunks[0]
+        else:
+            response = litellm.stream_chunk_builder(self.partial_response_chunks)
 
         try:
             # Use response_dict as a regular dictionary
@@ -3963,63 +3969,51 @@ class Coder(metaclass=UsageMeta):
                 if getattr(last_chunk, "usage", None):
                     response.usage = last_chunk.usage
 
-        # Collect provider-specific fields from chunks to preserve them
-        # We need to track both by ID (primary) and index (fallback) since
-        # early chunks might not have IDs established yet
-        provider_specific_fields_by_id = {}
-        provider_specific_fields_by_index = {}
-
+        # Collect message-level provider-specific fields (e.g. `reasoning_items`
+        # for reasoning models) from ALL chunks.  litellm's stream_chunk_builder()
+        # merges these with last-wins semantics for list fields, silently dropping
+        # every reasoning item except the final one.  Reasoning models depend on the
+        # full ordered item list being present in the assistant message so that
+        # exact-prefix prompt caching keeps working across turns, so we collect the
+        # fields ourselves and concatenate list-valued entries.
+        message_provider_specific_fields = {}
         for chunk in self.partial_response_chunks:
             try:
-                if chunk.choices and chunk.choices[0].delta and chunk.choices[0].delta.tool_calls:
-                    for tool_call in chunk.choices[0].delta.tool_calls:
-                        if (
-                            hasattr(tool_call, "provider_specific_fields")
-                            and tool_call.provider_specific_fields
-                        ):
-                            # Ensure provider_specific_fields is a dictionary
-                            psf = tool_call.provider_specific_fields
-                            if not isinstance(psf, dict):
-                                continue
-
-                            # Try to use ID first
-                            if hasattr(tool_call, "id") and tool_call.id:
-                                tool_id = tool_call.id
-                                if tool_id not in provider_specific_fields_by_id:
-                                    provider_specific_fields_by_id[tool_id] = {}
-                                # Merge provider-specific fields for this tool ID
-                                provider_specific_fields_by_id[tool_id].update(psf)
-                            # Also track by index as fallback
-                            elif hasattr(tool_call, "index"):
-                                tool_index = tool_call.index
-                                if tool_index not in provider_specific_fields_by_index:
-                                    provider_specific_fields_by_index[tool_index] = {}
-                                provider_specific_fields_by_index[tool_index].update(psf)
+                if chunk.choices and chunk.choices[0].delta:
+                    psf = getattr(chunk.choices[0].delta, "provider_specific_fields", None)
+                    if psf and isinstance(psf, dict):
+                        for key, value in psf.items():
+                            if isinstance(value, list):
+                                message_provider_specific_fields.setdefault(key, []).extend(value)
+                            elif value is not None:
+                                message_provider_specific_fields[key] = value
             except (AttributeError, IndexError):
                 continue
 
+        if message_provider_specific_fields:
+            message_psf = getattr(response.choices[0].message, "provider_specific_fields", None)
+            if not isinstance(message_psf, dict):
+                message_psf = {}
+            message_psf.update(message_provider_specific_fields)
+            response.choices[0].message.provider_specific_fields = message_psf
+
         try:
-            if response.choices[0].message.tool_calls:
-                for i, tool_call in enumerate(response.choices[0].message.tool_calls):
-                    # Add provider-specific fields if we collected any for this tool
-                    tool_id = tool_call.id
-
-                    # Try ID first
-                    if tool_id in provider_specific_fields_by_id:
-                        # Add provider-specific fields directly to the tool call object
-                        tool_call.provider_specific_fields = provider_specific_fields_by_id[tool_id]
-                    # Fall back to index
-                    elif i in provider_specific_fields_by_index:
-                        # Add provider-specific fields directly to the tool call object
-                        tool_call.provider_specific_fields = provider_specific_fields_by_index[i]
-
-                    # Only append to partial_response_tool_calls if it's empty
-                    if len(self.partial_response_tool_calls) == 0:
-                        self.partial_response_tool_calls.append(tool_call)
-
-                self.partial_response_function_call = (
-                    response.choices[0].message.tool_calls[0].function
-                )
+            message_tool_calls = response.choices[0].message.tool_calls
+            if message_tool_calls and len(message_tool_calls):
+                if self.stream:
+                    built_tool_calls = self._build_tool_calls_from_chunks()
+                    if built_tool_calls:
+                        response.choices[0].message.tool_calls = built_tool_calls
+                        self.partial_response_tool_calls = built_tool_calls
+                    else:
+                        # Fall back to litellm's merged list, keeping every call
+                        self.partial_response_tool_calls = list(message_tool_calls)
+                else:
+                    # Non-streaming: the single response chunk already carries the
+                    # full tool_calls list
+                    self.partial_response_tool_calls = list(message_tool_calls)
+
+                self.partial_response_function_call = self.partial_response_tool_calls[0].function
         except AttributeError as e:
             func_err = e
 
@@ -4075,6 +4069,85 @@ class Coder(metaclass=UsageMeta):
         self.partial_response_consolidated = (response, func_err, content_err)
         return response, func_err, content_err
 
+    def _build_tool_calls_from_chunks(self):
+        """Rebuild tool calls from the raw streaming chunks, keyed by delta index.
+
+        Streaming deltas for parallel tool calls arrive interleaved and may start
+        at any index (not necessarily 0).  Indexing into a dict by the delta's
+        tool-call ``index`` before converting it back to a list ensures every
+        parallel call is preserved, correctly ordered, and keeps its
+        provider-specific fields (e.g. thought signatures) attached.
+        """
+        from litellm.types.utils import ChatCompletionMessageToolCall, Function
+
+        tool_calls_dict = {}
+
+        for chunk in self.partial_response_chunks:
+            try:
+                if not (chunk.choices and chunk.choices[0].delta):
+                    continue
+
+                delta = chunk.choices[0].delta
+                for tool_call in delta.tool_calls or []:
+                    if tool_call is None:
+                        continue
+
+                    if nested.getter(tool_call, "function") is None:
+                        continue
+
+                    index = nested.getter(tool_call, "index")
+                    if index is None:
+                        index = len(tool_calls_dict)
+
+                    entry = tool_calls_dict.setdefault(
+                        index,
+                        {
+                            "id": None,
+                            "name": None,
+                            "type": "function",
+                            "arguments": [],
+                            "provider_specific_fields": {},
+                        },
+                    )
+
+                    entry["id"] = nested.getter(tool_call, "id") or entry["id"]
+                    entry["type"] = nested.getter(tool_call, "type") or entry["type"]
+                    entry["name"] = nested.getter(tool_call, "function.name") or entry["name"]
+
+                    arguments = nested.getter(tool_call, "function.arguments")
+                    if arguments:
+                        entry["arguments"].append(arguments)
+
+                    psf = nested.getter(tool_call, "provider_specific_fields")
+                    if not psf:
+                        psf = nested.getter(tool_call, "function.provider_specific_fields")
+                    if psf and isinstance(psf, dict):
+                        entry["provider_specific_fields"].update(psf)
+            except (AttributeError, IndexError):
+                continue
+
+        tool_calls = []
+        for index in sorted(tool_calls_dict.keys()):
+            data = tool_calls_dict[index]
+            if not (data["id"] and data["name"]):
+                continue
+
+            function = Function(
+                arguments="".join(data["arguments"]) or "{}",
+                name=data["name"],
+            )
+            params = {
+                "id": data["id"],
+                "function": function,
+                "type": data["type"] or "function",
+            }
+            if data["provider_specific_fields"]:
+                params["provider_specific_fields"] = data["provider_specific_fields"]
+
+            tool_calls.append(ChatCompletionMessageToolCall(**params))
+
+        return tool_calls
+
     def stream_wrapper(self, content, final):
         if not hasattr(self, "_streaming_buffer_length"):
             self._streaming_buffer_length = 0

{cecli_dev-1.0.5 → cecli_dev-1.1.0rc2}/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_dev-1.0.5 → cecli_dev-1.1.0rc2}/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_dev-1.0.5 → cecli_dev-1.1.0rc2}/cecli/helpers/io_proxy.py

@@ -51,6 +51,9 @@ class IOProxy(Generic[T]):
         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 @@ class IOProxy(Generic[T]):
         """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 @@ class IOProxy(Generic[T]):
 
     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_dev-1.1.0rc2/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_dev-1.1.0rc2/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_dev-1.1.0rc2/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_dev-1.1.0rc2/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_dev-1.1.0rc2/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

cecli_dev-1.1.0rc2/cecli/helpers/model_config/formatters/thinking.py

@@ -0,0 +1,108 @@
+"""Thinking formatters for the model config pipeline.
+
+These helpers rewrite the generic thinking shape onto the params litellm
+understands for a provider.  ``extra_params`` are the kwargs passed to
+``litellm.acompletion``, so a formatter exposes thinking 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
+  anthropic-style ``thinking`` param to ``thinkingBudget`` (Gemini 2.5) or
+  ``thinkingLevel`` (Gemini 3) and sets ``includeThoughts``.
+- Anthropic consumes the top-level ``thinking`` kwarg directly; riding inside
+  ``extra_body`` makes the API reject it as extra inputs.
+
+``set_thinking_tokens`` in :mod:`cecli.models` invokes the formatter chosen by
+the pipeline (``helpers.format_thinking``) after it has applied the generic
+shape.
+"""
+
+from __future__ import annotations
+
+from typing import Callable, Dict, Optional
+
+from ..identifiers import is_anthropic, is_claude_5_plus, is_gemini
+
+
+def format_thinking(provider: Optional[str], route: str, record: Optional[Dict]) -> Callable:
+    """Return the thinking 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_thinking_tokens`` keeps its default behavior.
+    """
+    if is_gemini(provider, route, record):
+        return gemini_thinking
+
+    if is_anthropic(provider, route, record):
+        if is_claude_5_plus(provider, route, record):
+            # Claude 5+ cannot use thinking.type.enabled; remove it entirely.
+            return anthropic_5_thinking
+
+        return anthropic_thinking
+
+    return noop
+
+
+def noop(extra_params: Dict) -> Dict:
+    """Default formatter: leave ``extra_params`` untouched."""
+    return extra_params
+
+
+def gemini_thinking(extra_params: Dict) -> Dict:
+    """Gemini models configure thinking via litellm's ``thinking`` kwarg.
+
+    litellm maps the top-level anthropic-style ``thinking`` param onto
+    Gemini's ``thinkingConfig`` (``thinkingBudget`` for Gemini 2.5,
+    ``thinkingLevel`` for Gemini 3) and sets ``includeThoughts``, so the
+    generic shape is lifted out of ``extra_body``.
+    """
+    return _lift_thinking(extra_params)
+
+
+def anthropic_thinking(extra_params: Dict) -> Dict:
+    """Anthropic (pre-5) models consume the top-level ``thinking`` kwarg.
+
+    litellm maps the top-level anthropic-style ``thinking`` param into the
+    request body, so it must not ride inside ``extra_body`` (which the
+    Anthropic API rejects as extra inputs).
+    """
+    params = _lift_thinking(extra_params)
+    params.pop("extra_body", None)
+    return params
+
+
+def anthropic_5_thinking(extra_params: Dict) -> Dict:
+    """Claude 5+ models cannot use ``thinking.type.enabled``.
+
+    Thinking is instead controlled via ``reasoning_effort`` (which litellm maps
+    to ``thinking.type.adaptive`` + ``output_config.effort``), so any thinking
+    block is removed entirely and ``extra_body`` is dropped.
+    """
+    extra_params.pop("thinking", None)
+    extra_body = extra_params.get("extra_body")
+
+    if isinstance(extra_body, dict):
+        extra_body.pop("thinking", None)
+
+    extra_params.pop("extra_body", None)
+    return extra_params
+
+
+def _lift_thinking(extra_params: Dict) -> Dict:
+    """Move the generic ``thinking`` shape out of ``extra_body`` to top level."""
+    extra_body = extra_params.get("extra_body")
+
+    if not isinstance(extra_body, dict):
+        return extra_params
+
+    thinking = extra_body.pop("thinking", None)
+
+    if thinking is not None:
+        extra_params["thinking"] = thinking
+
+    return extra_params