abstractcore 2.6.8__py3-none-any.whl → 2.9.0__py3-none-any.whl

abstractcore/tools/core.py

@@ -7,6 +7,43 @@ from dataclasses import dataclass, field
 from abc import ABC, abstractmethod
 
 
+_MAX_TOOL_DESCRIPTION_CHARS = 200
+_MAX_TOOL_WHEN_TO_USE_CHARS = 240
+_MAX_TOOL_EXAMPLES = 3
+
+
+def _first_non_empty_line(text: Optional[str]) -> str:
+    if not text:
+        return ""
+    for line in str(text).splitlines():
+        stripped = line.strip()
+        if stripped:
+            return stripped
+    return ""
+
+
+def _normalize_one_line(text: Optional[str]) -> str:
+    """Collapse whitespace (including newlines) into a single, prompt-friendly line."""
+    return " ".join(str(text or "").split()).strip()
+
+
+def _validate_tool_metadata(*, name: str, description: str, when_to_use: Optional[str], examples: List[Dict[str, Any]]) -> None:
+    if not description:
+        raise ValueError(f"Tool '{name}': description must be a non-empty string")
+    if len(description) > _MAX_TOOL_DESCRIPTION_CHARS:
+        raise ValueError(
+            f"Tool '{name}': description is too long ({len(description)} chars; max {_MAX_TOOL_DESCRIPTION_CHARS}). "
+            "Keep it to a single short sentence; put detailed guidance in `when_to_use` or docs."
+        )
+    if when_to_use is not None and len(when_to_use) > _MAX_TOOL_WHEN_TO_USE_CHARS:
+        raise ValueError(
+            f"Tool '{name}': when_to_use is too long ({len(when_to_use)} chars; max {_MAX_TOOL_WHEN_TO_USE_CHARS}). "
+            "Keep it to a single short sentence."
+        )
+    if len(examples) > _MAX_TOOL_EXAMPLES:
+        raise ValueError(f"Tool '{name}': too many examples ({len(examples)}; max {_MAX_TOOL_EXAMPLES}).")
+
+
 @dataclass
 class ToolDefinition:
     """Definition of a tool that can be called by LLM"""
@@ -20,6 +57,20 @@ class ToolDefinition:
     when_to_use: Optional[str] = None
     examples: List[Dict[str, Any]] = field(default_factory=list)
 
+    def __post_init__(self) -> None:
+        # Normalize to a single line for prompt-friendly catalogs.
+        self.name = str(self.name or "").strip()
+        self.description = _normalize_one_line(self.description)
+        self.when_to_use = _normalize_one_line(self.when_to_use) if self.when_to_use else None
+        self.tags = list(self.tags) if isinstance(self.tags, list) else []
+        self.examples = list(self.examples) if isinstance(self.examples, list) else []
+        _validate_tool_metadata(
+            name=self.name,
+            description=self.description,
+            when_to_use=self.when_to_use,
+            examples=self.examples,
+        )
+
     @classmethod
     def from_function(cls, func: Callable) -> 'ToolDefinition':
         """Create tool definition from a function"""
@@ -27,7 +78,11 @@ class ToolDefinition:
 
         # Extract function name and docstring
         name = func.__name__
-        description = func.__doc__ or "No description provided"
+        # Tool `description` must be short; use the first docstring line (not the whole docstring).
+        description = _first_non_empty_line(func.__doc__) or "No description provided"
+        description = _normalize_one_line(description)
+        if description != "No description provided":
+            _validate_tool_metadata(name=name, description=description, when_to_use=None, examples=[])
 
         # Extract parameters from function signature
         sig = inspect.signature(func)
@@ -59,17 +114,35 @@ class ToolDefinition:
 
     def to_dict(self) -> Dict[str, Any]:
         """Convert to dictionary format"""
-        result = {
-            "name": self.name,
-            "description": self.description,
-            "parameters": self.parameters
-        }
+        # Preserve a human/LLM-friendly field ordering (also improves UX when dumped as JSON).
+        result: Dict[str, Any] = {"name": self.name, "description": self.description}
+
+        if self.when_to_use:
+            result["when_to_use"] = self.when_to_use
+
+        result["parameters"] = self.parameters
+
+        # Expose required args explicitly for hosts that render ToolDefinitions directly
+        # (e.g. AbstractRuntime prompt payloads / debug traces). This is additive metadata:
+        # providers still rely on JSON Schema `required` when building native tool payloads.
+        try:
+            required: List[str] = []
+            for param_name, meta in (self.parameters or {}).items():
+                if not isinstance(param_name, str) or not param_name.strip():
+                    continue
+                # Convention in this repo: absence of `default` means "required".
+                if not isinstance(meta, dict) or "default" not in meta:
+                    required.append(param_name)
+            required.sort()
+            if required:
+                result["required_args"] = required
+        except Exception:
+            # Best-effort only; never break tool serialization.
+            pass
 
         # Include enhanced metadata if available
         if self.tags:
             result["tags"] = self.tags
-        if self.when_to_use:
-            result["when_to_use"] = self.when_to_use
         if self.examples:
             result["examples"] = self.examples
 
@@ -112,7 +185,8 @@ def tool(
     description: Optional[str] = None,
     tags: Optional[List[str]] = None,
     when_to_use: Optional[str] = None,
-    examples: Optional[List[Dict[str, Any]]] = None
+    examples: Optional[List[Dict[str, Any]]] = None,
+    hide_args: Optional[List[str]] = None,
 ):
     """
     Enhanced decorator to convert a function into a tool with rich metadata.
@@ -144,7 +218,8 @@ def tool(
     """
     def decorator(f):
         tool_name = name or f.__name__
-        tool_description = description or f.__doc__ or f"Execute {tool_name}"
+        tool_description = description or _first_non_empty_line(f.__doc__) or f"Execute {tool_name}"
+        tool_description = _normalize_one_line(tool_description)
 
         # Create tool definition from function and customize
         tool_def = ToolDefinition.from_function(f)
@@ -153,8 +228,29 @@ def tool(
 
         # Add enhanced metadata
         tool_def.tags = tags or []
-        tool_def.when_to_use = when_to_use
-        tool_def.examples = examples or []
+        tool_def.when_to_use = _normalize_one_line(when_to_use) if when_to_use else None
+        tool_def.examples = list(examples) if isinstance(examples, list) else []
+
+        # Optionally hide parameters from the exported schema (LLM-facing), while
+        # keeping them accepted by the underlying Python callable for backwards
+        # compatibility (e.g. legacy callers still passing deprecated kwargs).
+        hidden = [str(a).strip() for a in (hide_args or []) if str(a).strip()]
+        if hidden:
+            for arg in hidden:
+                if arg not in tool_def.parameters:
+                    continue
+                # Avoid hiding required args (no default), which would make the
+                # tool schema incomplete for tool-call generation.
+                if "default" not in tool_def.parameters.get(arg, {}):
+                    raise ValueError(f"Tool '{tool_def.name}': cannot hide required arg '{arg}'")
+                tool_def.parameters.pop(arg, None)
+
+        _validate_tool_metadata(
+            name=tool_def.name,
+            description=tool_def.description,
+            when_to_use=tool_def.when_to_use,
+            examples=tool_def.examples,
+        )
 
         # Attach tool definition to function for easy access
         f._tool_definition = tool_def
@@ -167,4 +263,4 @@ def tool(
         return decorator
     else:
         # Called without arguments: @tool
-        return decorator(func)
+        return decorator(func)

abstractcore/tools/handler.py

@@ -51,7 +51,10 @@ class UniversalToolHandler:
 
     def format_tools_prompt(
         self,
-        tools: List[Union[ToolDefinition, Callable, Dict[str, Any]]]
+        tools: List[Union[ToolDefinition, Callable, Dict[str, Any]]],
+        *,
+        include_tool_list: bool = True,
+        include_examples: bool = True,
     ) -> str:
         """
         Format tools into a system prompt for prompted models.
@@ -71,7 +74,12 @@ class UniversalToolHandler:
             return ""
 
         # Use architecture-specific formatting
-        return format_tool_prompt(tool_defs, self.model_name)
+        return format_tool_prompt(
+            tool_defs,
+            self.model_name,
+            include_tool_list=include_tool_list,
+            include_examples=include_examples,
+        )
 
     def prepare_tools_for_native(
         self,
@@ -243,6 +251,9 @@ class UniversalToolHandler:
                             logger.warning(f"Failed to parse tool arguments: {tool_call.arguments}")
                             tool_call.arguments = {}
 
+                    from .arg_canonicalizer import canonicalize_tool_arguments
+
+                    tool_call.arguments = canonicalize_tool_arguments(tool_call.name, tool_call.arguments)
                     tool_calls.append(tool_call)
             
             return ToolCallResponse(
@@ -271,6 +282,9 @@ class UniversalToolHandler:
                             logger.warning(f"Failed to parse tool arguments: {tool_call.arguments}")
                             tool_call.arguments = {}
 
+                    from .arg_canonicalizer import canonicalize_tool_arguments
+
+                    tool_call.arguments = canonicalize_tool_arguments(tool_call.name, tool_call.arguments)
                     tool_calls.append(tool_call)
 
             return ToolCallResponse(
@@ -305,4 +319,4 @@ def create_handler(model_name: str) -> UniversalToolHandler:
     Returns:
         Configured UniversalToolHandler
     """
-    return UniversalToolHandler(model_name)
+    return UniversalToolHandler(model_name)