rappel 0.10.0__py3-none-win_amd64.whl

rappel/ir_builder.py

@@ -0,0 +1,3945 @@
+"""
+IR Builder - Converts Python workflow AST to Rappel IR (ast.proto).
+
+This module parses Python workflow classes and produces the IR representation
+that can be sent to the Rust runtime for execution.
+
+The IR builder performs deep transformations to convert Python patterns into
+valid Rappel IR structures. Control flow bodies (try, for, while, if branches)
+are represented as full blocks, so they can contain arbitrary statements and
+`return` behaves consistently with Python (returns from the entry function end
+the workflow).
+
+Validation:
+The IR builder proactively detects unsupported Python patterns and raises
+UnsupportedPatternError with clear recommendations for how to rewrite the code.
+"""
+
+from __future__ import annotations
+
+import ast
+import copy
+import inspect
+import textwrap
+from dataclasses import dataclass
+from enum import EnumMeta
+from typing import TYPE_CHECKING, Any, Callable, Dict, List, Mapping, NoReturn, Optional, Set, Union
+
+from proto import ast_pb2 as ir
+from rappel.registry import registry
+
+
+class UnsupportedPatternError(Exception):
+    """Raised when the IR builder encounters an unsupported Python pattern.
+
+    This error includes a recommendation for how to rewrite the code to use
+    supported patterns.
+    """
+
+    def __init__(
+        self,
+        message: str,
+        recommendation: str,
+        line: Optional[int] = None,
+        col: Optional[int] = None,
+        filename: Optional[str] = None,
+    ):
+        self.message = message
+        self.recommendation = recommendation
+        self.line = line
+        self.col = col
+        self.filename = filename
+
+        location_parts: List[str] = []
+        if filename:
+            location_parts.append(filename)
+        if line:
+            location_parts.append(f"line {line}")
+        if col is not None:
+            location_parts.append(f"col {col}")
+        location = f" ({', '.join(location_parts)})" if location_parts else ""
+        full_message = f"{message}{location}\n\nRecommendation: {recommendation}"
+        super().__init__(full_message)
+
+
+# Recommendations for common unsupported patterns
+RECOMMENDATIONS = {
+    "constructor_return": (
+        "Returning a class constructor (like MyModel(...)) directly is not supported.\n"
+        "The workflow IR cannot serialize arbitrary object instantiation.\n\n"
+        "Use an @action to create the object:\n\n"
+        "    @action\n"
+        "    async def build_result(items: list, count: int) -> MyResult:\n"
+        "        return MyResult(items=items, count=count)\n\n"
+        "    # In workflow:\n"
+        "    return await build_result(items, count)"
+    ),
+    "constructor_assignment": (
+        "Assigning a class constructor result (like x = MyClass(...)) is not supported.\n"
+        "The workflow IR cannot serialize arbitrary object instantiation.\n\n"
+        "Use an @action to create the object:\n\n"
+        "    @action\n"
+        "    async def create_config(value: int) -> Config:\n"
+        "        return Config(value=value)\n\n"
+        "    # In workflow:\n"
+        "    config = await create_config(value)"
+    ),
+    "non_action_call": (
+        "Calling a function that is not decorated with @action is not supported.\n"
+        "Only @action decorated functions can be awaited in workflow code.\n\n"
+        "Add the @action decorator to your function:\n\n"
+        "    @action\n"
+        "    async def my_function(x: int) -> int:\n"
+        "        return x * 2"
+    ),
+    "sync_function_call": (
+        "Calling a synchronous function directly in workflow code is not supported.\n"
+        "All computation must happen inside @action decorated async functions.\n\n"
+        "Wrap your logic in an @action:\n\n"
+        "    @action\n"
+        "    async def compute(x: int) -> int:\n"
+        "        return some_sync_function(x)"
+    ),
+    "method_call_non_self": (
+        "Calling methods on objects other than 'self' is not supported in workflow code.\n"
+        "Use an @action to perform method calls:\n\n"
+        "    @action\n"
+        "    async def call_method(obj: MyClass) -> Result:\n"
+        "        return obj.some_method()"
+    ),
+    "builtin_call": (
+        "Calling built-in functions like len(), str(), int() directly is not supported.\n"
+        "Use an @action to perform these operations:\n\n"
+        "    @action\n"
+        "    async def get_length(items: list) -> int:\n"
+        "        return len(items)"
+    ),
+    "fstring": (
+        "F-strings are not supported in workflow code because they require "
+        "runtime string interpolation.\n"
+        "Use an @action to perform string formatting:\n\n"
+        "    @action\n"
+        "    async def format_message(value: int) -> str:\n"
+        "        return f'Result: {value}'"
+    ),
+    "delete": (
+        "The 'del' statement is not supported in workflow code.\n"
+        "Use an @action to perform mutations:\n\n"
+        "    @action\n"
+        "    async def remove_key(data: dict, key: str) -> dict:\n"
+        "        del data[key]\n"
+        "        return data"
+    ),
+    "while_loop": (
+        "While loops are not supported in workflow code because they can run "
+        "indefinitely.\n"
+        "Use a for loop with a fixed range, or restructure as recursive workflow calls."
+    ),
+    "with_statement": (
+        "Context managers (with statements) are not supported in workflow code.\n"
+        "Use an @action to handle resource management:\n\n"
+        "    @action\n"
+        "    async def read_file(path: str) -> str:\n"
+        "        with open(path) as f:\n"
+        "            return f.read()"
+    ),
+    "raise_statement": (
+        "The 'raise' statement is not supported directly in workflow code.\n"
+        "Use an @action that raises exceptions, or return error values."
+    ),
+    "assert_statement": (
+        "Assert statements are not supported in workflow code.\n"
+        "Use an @action for validation, or use if statements with explicit error handling."
+    ),
+    "lambda": (
+        "Lambda expressions are not supported in workflow code.\n"
+        "Use an @action to define the function logic."
+    ),
+    "list_comprehension": (
+        "List comprehensions are only supported when assigned directly to a variable "
+        "or inside asyncio.gather(*[...]).\n"
+        "For other cases, use a for loop or an @action."
+    ),
+    "dict_comprehension": (
+        "Dict comprehensions are only supported when assigned directly to a variable.\n"
+        "For other cases, use a for loop or an @action:\n\n"
+        "    @action\n"
+        "    async def build_dict(items: list) -> dict:\n"
+        "        return {k: v for k, v in items}"
+    ),
+    "set_comprehension": (
+        "Set comprehensions are not supported in workflow code.\nUse an @action to build sets."
+    ),
+    "generator": (
+        "Generator expressions are not supported in workflow code.\n"
+        "Use a list or an @action instead."
+    ),
+    "walrus": (
+        "The walrus operator (:=) is not supported in workflow code.\n"
+        "Use separate assignment statements instead."
+    ),
+    "match": (
+        "Match statements are not supported in workflow code.\nUse if/elif/else chains instead."
+    ),
+    "gather_return_exceptions": (
+        "asyncio.gather() must be called with return_exceptions=True so results and errors "
+        "are represented consistently in Rappel's IR.\n\n"
+        "Update your call:\n\n"
+        "    results = await asyncio.gather(task_a(), task_b(), return_exceptions=True)"
+    ),
+    "gather_variable_spread": (
+        "Spreading a variable in asyncio.gather() is not supported because it requires "
+        "data flow analysis to determine the contents.\n"
+        "Use a list comprehension directly in gather:\n\n"
+        "    # Instead of:\n"
+        "    tasks = []\n"
+        "    for i in range(count):\n"
+        "        tasks.append(process(value=i))\n"
+        "    results = await asyncio.gather(*tasks, return_exceptions=True)\n\n"
+        "    # Use:\n"
+        "    results = await asyncio.gather(*[process(value=i) for i in range(count)], "
+        "return_exceptions=True)"
+    ),
+    "for_loop_append_pattern": (
+        "Building a task list in a for loop then spreading in asyncio.gather() is not "
+        "supported.\n"
+        "Use a list comprehension directly in gather:\n\n"
+        "    # Instead of:\n"
+        "    tasks = []\n"
+        "    for i in range(count):\n"
+        "        tasks.append(process(value=i))\n"
+        "    results = await asyncio.gather(*tasks, return_exceptions=True)\n\n"
+        "    # Use:\n"
+        "    results = await asyncio.gather(*[process(value=i) for i in range(count)], "
+        "return_exceptions=True)"
+    ),
+    "global_statement": (
+        "Global statements are not supported in workflow code.\n"
+        "Use workflow state or pass values explicitly."
+    ),
+    "nonlocal_statement": (
+        "Nonlocal statements are not supported in workflow code.\n"
+        "Use explicit parameter passing instead."
+    ),
+    "import_statement": (
+        "Import statements inside workflow run() are not supported.\n"
+        "Place imports at the module level."
+    ),
+    "class_def": (
+        "Class definitions inside workflow run() are not supported.\n"
+        "Define classes at the module level."
+    ),
+    "function_def": (
+        "Nested function definitions inside workflow run() are not supported.\n"
+        "Define functions at the module level or use @action."
+    ),
+    "yield_statement": (
+        "Yield statements are not supported in workflow code.\n"
+        "Workflows must return a complete result, not generate values incrementally."
+    ),
+    "continue_statement": (
+        "Continue statements are not supported in workflow code.\n"
+        "Restructure your loop using if/else to skip iterations."
+    ),
+    "unsupported_statement": (
+        "This statement type is not supported in workflow code.\n"
+        "Move the logic into an @action or rewrite using supported statements."
+    ),
+    "unsupported_expression": (
+        "This expression type is not supported in workflow code.\n"
+        "Move the logic into an @action or rewrite using supported expressions."
+    ),
+    "unsupported_literal": (
+        "This literal type is not supported in workflow code.\n"
+        "Convert the value to a supported literal type inside an @action."
+    ),
+}
+
+GLOBAL_FUNCTIONS = {
+    "enumerate": ir.GlobalFunction.GLOBAL_FUNCTION_ENUMERATE,
+    "isexception": ir.GlobalFunction.GLOBAL_FUNCTION_ISEXCEPTION,
+    "len": ir.GlobalFunction.GLOBAL_FUNCTION_LEN,
+    "range": ir.GlobalFunction.GLOBAL_FUNCTION_RANGE,
+}
+ALLOWED_SYNC_FUNCTIONS = set(GLOBAL_FUNCTIONS)
+
+_CURRENT_ACTION_NAMES: set[str] = set()
+
+
+if TYPE_CHECKING:
+    from .workflow import Workflow
+
+
+@dataclass
+class ActionDefinition:
+    """Definition of an action function."""
+
+    action_name: str
+    module_name: Optional[str]
+    signature: inspect.Signature
+
+
+@dataclass
+class FunctionParameter:
+    """A function parameter with optional default value."""
+
+    name: str
+    has_default: bool
+    default_value: Any = None  # The actual Python value if has_default is True
+
+
+@dataclass
+class FunctionSignatureInfo:
+    """Signature info for a workflow function, used to fill in default arguments."""
+
+    parameters: List[FunctionParameter]
+
+
+@dataclass
+class ModuleContext:
+    """Cached IRBuilder context derived from a module."""
+
+    action_defs: Dict[str, ActionDefinition]
+    imported_names: Dict[str, "ImportedName"]
+    module_functions: Set[str]
+    model_defs: Dict[str, "ModelDefinition"]
+
+
+@dataclass
+class TransformContext:
+    """Context for IR transformations."""
+
+    # Counter for generating unique function names
+    implicit_fn_counter: int = 0
+    # Implicit functions generated during transformation
+    implicit_functions: List[ir.FunctionDef] = None  # type: ignore
+    # Function signatures for workflow methods (used to fill in default arguments)
+    function_signatures: Dict[str, FunctionSignatureInfo] = None  # type: ignore
+
+    def __post_init__(self) -> None:
+        if self.implicit_functions is None:
+            self.implicit_functions = []
+        if self.function_signatures is None:
+            self.function_signatures = {}
+
+    def next_implicit_fn_name(self, prefix: str = "implicit") -> str:
+        """Generate a unique implicit function name."""
+        self.implicit_fn_counter += 1
+        return f"__{prefix}_{self.implicit_fn_counter}__"
+
+
+def build_workflow_ir(workflow_cls: type["Workflow"]) -> ir.Program:
+    """Build an IR Program from a workflow class.
+
+    Args:
+        workflow_cls: The workflow class to convert.
+
+    Returns:
+        An IR Program proto message.
+    """
+    original_run = getattr(workflow_cls, "__workflow_run_impl__", None)
+    if original_run is None:
+        original_run = workflow_cls.__dict__.get("run")
+    if original_run is None:
+        raise ValueError(f"workflow {workflow_cls!r} missing run implementation")
+
+    module = inspect.getmodule(original_run)
+    if module is None:
+        raise ValueError(f"unable to locate module for workflow {workflow_cls!r}")
+
+    module_contexts: Dict[str, ModuleContext] = {}
+
+    def get_module_context(target_module: Any) -> ModuleContext:
+        module_name = target_module.__name__
+        if module_name not in module_contexts:
+            module_contexts[module_name] = ModuleContext(
+                action_defs=_discover_action_names(target_module),
+                imported_names=_discover_module_imports(target_module),
+                module_functions=_discover_module_functions(target_module),
+                model_defs=_discover_model_definitions(target_module),
+            )
+        return module_contexts[module_name]
+
+    # Build the IR with transformation context
+    ctx = TransformContext()
+    program = ir.Program()
+    function_defs: Dict[str, ir.FunctionDef] = {}
+
+    # Extract instance attributes from __init__ for policy resolution
+    instance_attrs = _extract_instance_attrs(workflow_cls)
+
+    def parse_function(fn: Any) -> tuple[ast.AST, Optional[str], int]:
+        source_lines, start_line = inspect.getsourcelines(fn)
+        function_source = textwrap.dedent("".join(source_lines))
+        filename = inspect.getsourcefile(fn)
+        if filename is None:
+            filename = inspect.getfile(fn)
+        return ast.parse(function_source, filename=filename or "<unknown>"), filename, start_line
+
+    def _with_source_location(
+        err: UnsupportedPatternError,
+        filename: Optional[str],
+        start_line: int,
+    ) -> UnsupportedPatternError:
+        line = err.line
+        col = err.col
+        if line is not None:
+            line = start_line + line - 1
+        if col is not None:
+            col = col + 1
+        return UnsupportedPatternError(
+            err.message,
+            err.recommendation,
+            line=line,
+            col=col,
+            filename=filename,
+        )
+
+    def add_function_def(
+        fn: Any,
+        fn_tree: ast.AST,
+        filename: Optional[str],
+        start_line: int,
+        override_name: Optional[str] = None,
+    ) -> None:
+        global _CURRENT_ACTION_NAMES
+        fn_module = inspect.getmodule(fn)
+        if fn_module is None:
+            raise ValueError(f"unable to locate module for function {fn!r}")
+
+        ctx_data = get_module_context(fn_module)
+        _CURRENT_ACTION_NAMES = set(ctx_data.action_defs.keys())
+        builder = IRBuilder(
+            ctx_data.action_defs,
+            ctx,
+            ctx_data.imported_names,
+            ctx_data.module_functions,
+            ctx_data.model_defs,
+            fn_module.__dict__,
+            instance_attrs,
+        )
+        try:
+            builder.visit(fn_tree)
+        except UnsupportedPatternError as err:
+            raise _with_source_location(err, filename, start_line) from err
+        if builder.function_def:
+            if override_name:
+                builder.function_def.name = override_name
+            function_defs[builder.function_def.name] = builder.function_def
+
+    # Discover all reachable helper methods first so we can pre-collect their signatures.
+    # This is needed because when we process a function that calls another function,
+    # we need to know the callee's signature to fill in default arguments.
+    run_tree, run_filename, run_start_line = parse_function(original_run)
+
+    # Collect all reachable methods and their trees
+    methods_to_process: List[tuple[Any, ast.AST, Optional[str], int, Optional[str]]] = [
+        (original_run, run_tree, run_filename, run_start_line, "main")
+    ]
+    pending = list(_collect_self_method_calls(run_tree))
+    visited: Set[str] = set()
+    skip_methods = {"run_action"}
+
+    while pending:
+        method_name = pending.pop()
+        if method_name in visited or method_name == "run" or method_name in skip_methods:
+            continue
+        visited.add(method_name)
+
+        method = _find_workflow_method(workflow_cls, method_name)
+        if method is None:
+            continue
+
+        method_tree, method_filename, method_start_line = parse_function(method)
+        methods_to_process.append((method, method_tree, method_filename, method_start_line, None))
+        pending.extend(_collect_self_method_calls(method_tree))
+
+    # Pre-collect signatures for all methods before processing IR.
+    # This ensures that when we encounter a function call, we can fill in default args.
+    for fn, _fn_tree, _filename, _start_line, override_name in methods_to_process:
+        fn_name = override_name if override_name else fn.__name__
+        sig = inspect.signature(fn)
+        params: List[FunctionParameter] = []
+        for param_name, param in sig.parameters.items():
+            if param_name == "self":
+                continue
+            has_default = param.default is not inspect.Parameter.empty
+            default_value = param.default if has_default else None
+            params.append(
+                FunctionParameter(
+                    name=param_name,
+                    has_default=has_default,
+                    default_value=default_value,
+                )
+            )
+        ctx.function_signatures[fn_name] = FunctionSignatureInfo(parameters=params)
+
+    # Now process all functions with signatures already available
+    for fn, fn_tree, filename, start_line, override_name in methods_to_process:
+        add_function_def(fn, fn_tree, filename, start_line, override_name)
+
+    # Add implicit functions first (they may be called by the main function)
+    for implicit_fn in ctx.implicit_functions:
+        program.functions.append(implicit_fn)
+
+    # Add all function definitions (run + reachable helper methods)
+    for fn_def in function_defs.values():
+        program.functions.append(fn_def)
+
+    global _CURRENT_ACTION_NAMES
+    _CURRENT_ACTION_NAMES = set()
+
+    return program
+
+
+def _collect_self_method_calls(tree: ast.AST) -> Set[str]:
+    """Collect self.method(...) call names from a parsed function AST."""
+    calls: Set[str] = set()
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Call):
+            continue
+        func = node.func
+        if isinstance(func, ast.Attribute) and isinstance(func.value, ast.Name):
+            if func.value.id == "self":
+                calls.add(func.attr)
+    return calls
+
+
+def _find_workflow_method(workflow_cls: type["Workflow"], name: str) -> Optional[Any]:
+    """Find a workflow method by name across the class MRO."""
+    for base in workflow_cls.__mro__:
+        if name not in base.__dict__:
+            continue
+        value = base.__dict__[name]
+        if isinstance(value, staticmethod) or isinstance(value, classmethod):
+            return value.__func__
+        if inspect.isfunction(value):
+            return value
+    return None
+
+
+def _extract_instance_attrs(workflow_cls: type["Workflow"]) -> Dict[str, ast.expr]:
+    """Extract self.attr = value assignments from the workflow's __init__ method.
+
+    Parses the __init__ method to find assignments like:
+        self.retry_policy = RetryPolicy(attempts=3)
+        self.timeout = 30
+
+    Returns a dict mapping attribute names to their AST value nodes.
+    """
+    init_method = _find_workflow_method(workflow_cls, "__init__")
+    if init_method is None:
+        return {}
+
+    try:
+        source_lines, _ = inspect.getsourcelines(init_method)
+        source = textwrap.dedent("".join(source_lines))
+        tree = ast.parse(source)
+    except (OSError, TypeError, SyntaxError):
+        return {}
+
+    attrs: Dict[str, ast.expr] = {}
+
+    # Walk the __init__ body looking for self.attr = value assignments
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Assign):
+            continue
+        # Only handle single-target assignments
+        if len(node.targets) != 1:
+            continue
+        target = node.targets[0]
+        # Check for self.attr pattern
+        if (
+            isinstance(target, ast.Attribute)
+            and isinstance(target.value, ast.Name)
+            and target.value.id == "self"
+        ):
+            attrs[target.attr] = node.value
+
+    return attrs
+
+
+def _discover_action_names(module: Any) -> Dict[str, ActionDefinition]:
+    """Discover all @action decorated functions in a module."""
+    names: Dict[str, ActionDefinition] = {}
+    for attr_name in dir(module):
+        attr = getattr(module, attr_name)
+        action_name = getattr(attr, "__rappel_action_name__", None)
+        action_module = getattr(attr, "__rappel_action_module__", None)
+        if callable(attr) and action_name:
+            signature = inspect.signature(attr)
+            names[attr_name] = ActionDefinition(
+                action_name=action_name,
+                module_name=action_module or module.__name__,
+                signature=signature,
+            )
+    for entry in registry.entries():
+        if entry.module != module.__name__:
+            continue
+        func_name = entry.func.__name__
+        if func_name in names:
+            continue
+        signature = inspect.signature(entry.func)
+        names[func_name] = ActionDefinition(
+            action_name=entry.name,
+            module_name=entry.module,
+            signature=signature,
+        )
+    return names
+
+
+def _discover_module_functions(module: Any) -> Set[str]:
+    """Discover all async function names defined in a module.
+
+    This is used to detect when users await functions in the same module
+    that are NOT decorated with @action.
+    """
+    function_names: Set[str] = set()
+    for attr_name in dir(module):
+        try:
+            attr = getattr(module, attr_name)
+        except AttributeError:
+            continue
+
+        # Only include functions defined in THIS module (not imported)
+        if not callable(attr):
+            continue
+        if not inspect.iscoroutinefunction(attr):
+            continue
+
+        # Check if the function is defined in this module
+        func_module = getattr(attr, "__module__", None)
+        if func_module == module.__name__:
+            function_names.add(attr_name)
+
+    return function_names
+
+
+@dataclass
+class ImportedName:
+    """Tracks an imported name and its source module."""
+
+    local_name: str  # Name used in code (e.g., "sleep")
+    module: str  # Source module (e.g., "asyncio")
+    original_name: str  # Original name in source module (e.g., "sleep")
+
+
+@dataclass
+class ModelFieldDefinition:
+    """Definition of a field in a Pydantic model or dataclass."""
+
+    name: str
+    has_default: bool
+    default_value: Any = None  # Only set if has_default is True
+
+
+@dataclass
+class ModelDefinition:
+    """Definition of a Pydantic model or dataclass that can be used in workflows.
+
+    These are data classes that can be instantiated in workflow code and will
+    be converted to dictionary expressions in the IR.
+    """
+
+    class_name: str
+    fields: Dict[str, ModelFieldDefinition]
+    is_pydantic: bool  # True for Pydantic models, False for dataclasses
+
+
+def _is_simple_pydantic_model(cls: type) -> bool:
+    """Check if a class is a simple Pydantic model without custom validators.
+
+    A simple Pydantic model:
+    - Inherits from pydantic.BaseModel
+    - Has no field_validator or model_validator decorators
+    - Has no custom __init__ method
+
+    Returns False if pydantic is not installed or cls is not a Pydantic model.
+    """
+    try:
+        from pydantic import BaseModel
+    except ImportError:
+        return False
+
+    if not isinstance(cls, type) or not issubclass(cls, BaseModel):
+        return False
+
+    # Check for validators - Pydantic v2 uses __pydantic_decorators__
+    decorators = getattr(cls, "__pydantic_decorators__", None)
+    if decorators is not None:
+        # Check for field validators
+        if hasattr(decorators, "field_validators") and decorators.field_validators:
+            return False
+        # Check for model validators
+        if hasattr(decorators, "model_validators") and decorators.model_validators:
+            return False
+
+    # Check for custom __init__ (not the one from BaseModel)
+    if "__init__" in cls.__dict__:
+        return False
+
+    return True
+
+
+def _is_simple_dataclass(cls: type) -> bool:
+    """Check if a class is a simple dataclass without custom logic.
+
+    A simple dataclass:
+    - Is decorated with @dataclass
+    - Has no custom __init__ method (uses the generated one)
+    - Has no __post_init__ method
+
+    Returns False if cls is not a dataclass.
+    """
+    import dataclasses
+
+    if not dataclasses.is_dataclass(cls):
+        return False
+
+    # Check for __post_init__ which could have custom logic
+    if hasattr(cls, "__post_init__") and "__post_init__" in cls.__dict__:
+        return False
+
+    # Dataclasses generate __init__, so check if there's a custom one
+    # that overrides it (unlikely but possible)
+    # The dataclass decorator sets __init__ so we can't easily detect override
+    # We'll trust that dataclasses without __post_init__ are simple
+
+    return True
+
+
+def _get_pydantic_model_fields(cls: type) -> Dict[str, ModelFieldDefinition]:
+    """Extract field definitions from a Pydantic model."""
+    try:
+        from pydantic import BaseModel
+        from pydantic.fields import FieldInfo
+    except ImportError:
+        return {}
+
+    if not issubclass(cls, BaseModel):
+        return {}
+
+    fields: Dict[str, ModelFieldDefinition] = {}
+
+    # Pydantic v2 uses model_fields
+    model_fields = getattr(cls, "model_fields", {})
+    for field_name, field_info in model_fields.items():
+        has_default = False
+        default_value = None
+
+        if isinstance(field_info, FieldInfo):
+            # Check if field has a default value
+            # PydanticUndefined means no default
+            from pydantic_core import PydanticUndefined
+
+            if field_info.default is not PydanticUndefined:
+                has_default = True
+                default_value = field_info.default
+            elif field_info.default_factory is not None:
+                # We can't serialize factory functions, so treat as no default
+                has_default = False
+
+        fields[field_name] = ModelFieldDefinition(
+            name=field_name,
+            has_default=has_default,
+            default_value=default_value,
+        )
+
+    return fields
+
+
+def _get_dataclass_fields(cls: type) -> Dict[str, ModelFieldDefinition]:
+    """Extract field definitions from a dataclass."""
+    import dataclasses
+
+    if not dataclasses.is_dataclass(cls):
+        return {}
+
+    fields: Dict[str, ModelFieldDefinition] = {}
+    for field in dataclasses.fields(cls):
+        has_default = False
+        default_value = None
+
+        if field.default is not dataclasses.MISSING:
+            has_default = True
+            default_value = field.default
+        elif field.default_factory is not dataclasses.MISSING:
+            # We can't serialize factory functions, so treat as no default
+            has_default = False
+
+        fields[field.name] = ModelFieldDefinition(
+            name=field.name,
+            has_default=has_default,
+            default_value=default_value,
+        )
+
+    return fields
+
+
+def _discover_model_definitions(module: Any) -> Dict[str, ModelDefinition]:
+    """Discover all Pydantic models and dataclasses that can be used in workflows.
+
+    Only discovers "simple" models without custom validators or __post_init__.
+    """
+    models: Dict[str, ModelDefinition] = {}
+
+    for attr_name in dir(module):
+        try:
+            attr = getattr(module, attr_name)
+        except AttributeError:
+            continue
+
+        if not isinstance(attr, type):
+            continue
+
+        # Check if this class is defined in this module or imported
+        # We want to include both, as models might be imported
+        if _is_simple_pydantic_model(attr):
+            fields = _get_pydantic_model_fields(attr)
+            models[attr_name] = ModelDefinition(
+                class_name=attr_name,
+                fields=fields,
+                is_pydantic=True,
+            )
+        elif _is_simple_dataclass(attr):
+            fields = _get_dataclass_fields(attr)
+            models[attr_name] = ModelDefinition(
+                class_name=attr_name,
+                fields=fields,
+                is_pydantic=False,
+            )
+
+    return models
+
+
+def _discover_module_imports(module: Any) -> Dict[str, ImportedName]:
+    """Discover imports in a module by parsing its source.
+
+    Tracks imports like:
+    - from asyncio import sleep  -> {"sleep": ImportedName("sleep", "asyncio", "sleep")}
+    - from asyncio import sleep as s -> {"s": ImportedName("s", "asyncio", "sleep")}
+    """
+    imported: Dict[str, ImportedName] = {}
+
+    try:
+        source = inspect.getsource(module)
+        tree = ast.parse(source)
+    except (OSError, TypeError):
+        # Can't get source (e.g., built-in module)
+        return imported
+
+    for node in ast.walk(tree):
+        if isinstance(node, ast.ImportFrom) and node.module:
+            for alias in node.names:
+                local_name = alias.asname if alias.asname else alias.name
+                imported[local_name] = ImportedName(
+                    local_name=local_name,
+                    module=node.module,
+                    original_name=alias.name,
+                )
+
+    return imported
+
+
+class IRBuilder(ast.NodeVisitor):
+    """Builds IR from Python AST with deep transformations."""
+
+    def __init__(
+        self,
+        action_defs: Dict[str, ActionDefinition],
+        ctx: TransformContext,
+        imported_names: Optional[Dict[str, ImportedName]] = None,
+        module_functions: Optional[Set[str]] = None,
+        model_defs: Optional[Dict[str, ModelDefinition]] = None,
+        module_globals: Optional[Mapping[str, Any]] = None,
+        instance_attrs: Optional[Dict[str, ast.expr]] = None,
+    ):
+        self._action_defs = action_defs
+        self._ctx = ctx
+        self._imported_names = imported_names or {}
+        self._module_functions = module_functions or set()
+        self._model_defs = model_defs or {}
+        self._module_globals = module_globals or {}
+        self._instance_attrs = instance_attrs or {}
+        self.function_def: Optional[ir.FunctionDef] = None
+        self._statements: List[ir.Statement] = []
+
+    def visit_FunctionDef(self, node: ast.FunctionDef) -> Any:
+        """Visit a function definition (the workflow's run method)."""
+        inputs = self._collect_function_inputs(node)
+
+        # Create the function definition
+        self.function_def = ir.FunctionDef(
+            name=node.name,
+            io=ir.IoDecl(inputs=inputs, outputs=[]),
+            span=_make_span(node),
+        )
+
+        # Visit the body - _visit_statement now returns a list
+        self._statements = []
+        for stmt in node.body:
+            ir_stmts = self._visit_statement(stmt)
+            self._statements.extend(ir_stmts)
+
+        # Set the body
+        self.function_def.body.CopyFrom(ir.Block(statements=self._statements))
+
+    def visit_AsyncFunctionDef(self, node: ast.AsyncFunctionDef) -> Any:
+        """Visit an async function definition (the workflow's run method)."""
+        # Handle async the same way as sync for IR building
+        inputs = self._collect_function_inputs(node)
+
+        self.function_def = ir.FunctionDef(
+            name=node.name,
+            io=ir.IoDecl(inputs=inputs, outputs=[]),
+            span=_make_span(node),
+        )
+
+        self._statements = []
+        for stmt in node.body:
+            ir_stmts = self._visit_statement(stmt)
+            self._statements.extend(ir_stmts)
+
+        self.function_def.body.CopyFrom(ir.Block(statements=self._statements))
+
+    def _visit_statement(self, node: ast.stmt) -> List[ir.Statement]:
+        """Convert a Python statement to IR Statement(s).
+
+        Returns a list because some transformations (like try block hoisting)
+        may expand a single Python statement into multiple IR statements.
+
+        Raises UnsupportedPatternError for unsupported statement types.
+        """
+        if isinstance(node, ast.Assign):
+            dict_expanded = self._expand_dict_comprehension_assignment(node)
+            if dict_expanded is not None:
+                return dict_expanded
+            expanded = self._expand_list_comprehension_assignment(node)
+            if expanded is not None:
+                return expanded
+            result = self._visit_assign(node)
+            return [result] if result else []
+        elif isinstance(node, ast.AnnAssign):
+            result = self._visit_ann_assign(node)
+            return [result] if result else []
+        elif isinstance(node, ast.Expr):
+            result = self._visit_expr_stmt(node)
+            return [result] if result else []
+        elif isinstance(node, ast.For):
+            return self._visit_for(node)
+        elif isinstance(node, ast.While):
+            return self._visit_while(node)
+        elif isinstance(node, ast.If):
+            return self._visit_if(node)
+        elif isinstance(node, ast.Try):
+            return self._visit_try(node)
+        elif isinstance(node, ast.Return):
+            return self._visit_return(node)
+        elif isinstance(node, ast.AugAssign):
+            return self._visit_aug_assign(node)
+        elif isinstance(node, ast.Pass):
+            # Pass statements are fine, they just don't produce IR
+            return []
+        elif isinstance(node, ast.Break):
+            return self._visit_break(node)
+        elif isinstance(node, ast.Continue):
+            return self._visit_continue(node)
+
+        # Check for unsupported statement types - this MUST raise for any
+        # unhandled statement to avoid silently dropping code
+        self._check_unsupported_statement(node)
+
+    def _check_unsupported_statement(self, node: ast.stmt) -> NoReturn:
+        """Check for unsupported statement types and raise descriptive errors.
+
+        This function ALWAYS raises an exception - it never returns normally.
+        Any statement type that reaches this function is either explicitly
+        unsupported (with a specific error message) or unhandled (with a
+        generic catch-all error). This ensures we never silently drop code.
+        """
+        line = getattr(node, "lineno", None)
+        col = getattr(node, "col_offset", None)
+
+        if isinstance(node, (ast.With, ast.AsyncWith)):
+            raise UnsupportedPatternError(
+                "Context managers (with statements) are not supported",
+                RECOMMENDATIONS["with_statement"],
+                line=line,
+                col=col,
+            )
+        elif isinstance(node, ast.Raise):
+            raise UnsupportedPatternError(
+                "The 'raise' statement is not supported",
+                RECOMMENDATIONS["raise_statement"],
+                line=line,
+                col=col,
+            )
+        elif isinstance(node, ast.Assert):
+            raise UnsupportedPatternError(
+                "Assert statements are not supported",
+                RECOMMENDATIONS["assert_statement"],
+                line=line,
+                col=col,
+            )
+        elif isinstance(node, ast.Delete):
+            raise UnsupportedPatternError(
+                "The 'del' statement is not supported",
+                RECOMMENDATIONS["delete"],
+                line=line,
+                col=col,
+            )
+        elif isinstance(node, ast.Global):
+            raise UnsupportedPatternError(
+                "Global statements are not supported",
+                RECOMMENDATIONS["global_statement"],
+                line=line,
+                col=col,
+            )
+        elif isinstance(node, ast.Nonlocal):
+            raise UnsupportedPatternError(
+                "Nonlocal statements are not supported",
+                RECOMMENDATIONS["nonlocal_statement"],
+                line=line,
+                col=col,
+            )
+        elif isinstance(node, (ast.Import, ast.ImportFrom)):
+            raise UnsupportedPatternError(
+                "Import statements inside workflow run() are not supported",
+                RECOMMENDATIONS["import_statement"],
+                line=line,
+                col=col,
+            )
+        elif isinstance(node, ast.ClassDef):
+            raise UnsupportedPatternError(
+                "Class definitions inside workflow run() are not supported",
+                RECOMMENDATIONS["class_def"],
+                line=line,
+                col=col,
+            )
+        elif isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
+            raise UnsupportedPatternError(
+                "Nested function definitions are not supported",
+                RECOMMENDATIONS["function_def"],
+                line=line,
+                col=col,
+            )
+        elif hasattr(ast, "Match") and isinstance(node, ast.Match):
+            raise UnsupportedPatternError(
+                "Match statements are not supported",
+                RECOMMENDATIONS["match"],
+                line=line,
+                col=col,
+            )
+        else:
+            # Catch-all for any unhandled statement types.
+            # This is critical to avoid silently dropping code.
+            stmt_type = type(node).__name__
+            raise UnsupportedPatternError(
+                f"Unhandled statement type: {stmt_type}",
+                RECOMMENDATIONS["unsupported_statement"],
+                line=line,
+                col=col,
+            )
+
+    def _expand_list_comprehension_assignment(
+        self, node: ast.Assign
+    ) -> Optional[List[ir.Statement]]:
+        """Expand a list comprehension assignment into loop-based statements.
+
+        Example:
+            active_users = [user for user in users if user.active]
+
+        Becomes:
+            active_users = []
+            for user in users:
+                if user.active:
+                    active_users = active_users + [user]
+        """
+        if not isinstance(node.value, ast.ListComp):
+            return None
+
+        if len(node.targets) != 1 or not isinstance(node.targets[0], ast.Name):
+            line = getattr(node, "lineno", None)
+            col = getattr(node, "col_offset", None)
+            raise UnsupportedPatternError(
+                "List comprehension assignments must target a single variable",
+                "Assign the comprehension to a simple variable like `results = [x for x in items]`",
+                line=line,
+                col=col,
+            )
+
+        listcomp = node.value
+        if len(listcomp.generators) != 1:
+            line = getattr(listcomp, "lineno", None)
+            col = getattr(listcomp, "col_offset", None)
+            raise UnsupportedPatternError(
+                "List comprehensions with multiple generators are not supported",
+                "Use nested for loops instead of combining multiple generators in one comprehension",
+                line=line,
+                col=col,
+            )
+
+        gen = listcomp.generators[0]
+        if gen.is_async:
+            line = getattr(listcomp, "lineno", None)
+            col = getattr(listcomp, "col_offset", None)
+            raise UnsupportedPatternError(
+                "Async list comprehensions are not supported",
+                "Rewrite using an explicit async for loop",
+                line=line,
+                col=col,
+            )
+
+        target_name = node.targets[0].id
+
+        # Initialize the accumulator list: active_users = []
+        init_assign_ast = ast.Assign(
+            targets=[ast.Name(id=target_name, ctx=ast.Store())],
+            value=ast.List(elts=[], ctx=ast.Load()),
+            type_comment=None,
+        )
+        ast.copy_location(init_assign_ast, node)
+        ast.fix_missing_locations(init_assign_ast)
+
+        def _make_append_assignment(value_expr: ast.expr) -> ast.Assign:
+            append_assign = ast.Assign(
+                targets=[ast.Name(id=target_name, ctx=ast.Store())],
+                value=ast.BinOp(
+                    left=ast.Name(id=target_name, ctx=ast.Load()),
+                    op=ast.Add(),
+                    right=ast.List(elts=[copy.deepcopy(value_expr)], ctx=ast.Load()),
+                ),
+                type_comment=None,
+            )
+            ast.copy_location(append_assign, node.value)
+            ast.fix_missing_locations(append_assign)
+            return append_assign
+
+        append_statements: List[ast.stmt] = []
+        if isinstance(listcomp.elt, ast.IfExp):
+            then_assign = _make_append_assignment(listcomp.elt.body)
+            else_assign = _make_append_assignment(listcomp.elt.orelse)
+            branch_if = ast.If(
+                test=copy.deepcopy(listcomp.elt.test),
+                body=[then_assign],
+                orelse=[else_assign],
+            )
+            ast.copy_location(branch_if, listcomp.elt)
+            ast.fix_missing_locations(branch_if)
+            append_statements.append(branch_if)
+        else:
+            append_statements.append(_make_append_assignment(listcomp.elt))
+
+        loop_body: List[ast.stmt] = append_statements
+        if gen.ifs:
+            condition: ast.expr
+            if len(gen.ifs) == 1:
+                condition = copy.deepcopy(gen.ifs[0])
+            else:
+                condition = ast.BoolOp(op=ast.And(), values=[copy.deepcopy(iff) for iff in gen.ifs])
+                ast.copy_location(condition, gen.ifs[0])
+            if_stmt = ast.If(test=condition, body=append_statements, orelse=[])
+            ast.copy_location(if_stmt, gen.ifs[0])
+            ast.fix_missing_locations(if_stmt)
+            loop_body = [if_stmt]
+
+        loop_ast = ast.For(
+            target=copy.deepcopy(gen.target),
+            iter=copy.deepcopy(gen.iter),
+            body=loop_body,
+            orelse=[],
+            type_comment=None,
+        )
+        ast.copy_location(loop_ast, node)
+        ast.fix_missing_locations(loop_ast)
+
+        statements: List[ir.Statement] = []
+        init_stmt = self._visit_assign(init_assign_ast)
+        if init_stmt:
+            statements.append(init_stmt)
+        statements.extend(self._visit_for(loop_ast))
+
+        return statements
+
+    def _expand_dict_comprehension_assignment(
+        self, node: ast.Assign
+    ) -> Optional[List[ir.Statement]]:
+        """Expand a dict comprehension assignment into loop-based statements."""
+        if not isinstance(node.value, ast.DictComp):
+            return None
+
+        if len(node.targets) != 1 or not isinstance(node.targets[0], ast.Name):
+            line = getattr(node, "lineno", None)
+            col = getattr(node, "col_offset", None)
+            raise UnsupportedPatternError(
+                "Dict comprehension assignments must target a single variable",
+                "Assign the comprehension to a simple variable like `result = {k: v for k, v in pairs}`",
+                line=line,
+                col=col,
+            )
+
+        dictcomp = node.value
+        if len(dictcomp.generators) != 1:
+            line = getattr(dictcomp, "lineno", None)
+            col = getattr(dictcomp, "col_offset", None)
+            raise UnsupportedPatternError(
+                "Dict comprehensions with multiple generators are not supported",
+                "Use nested for loops instead of combining multiple generators in one comprehension",
+                line=line,
+                col=col,
+            )
+
+        gen = dictcomp.generators[0]
+        if gen.is_async:
+            line = getattr(dictcomp, "lineno", None)
+            col = getattr(dictcomp, "col_offset", None)
+            raise UnsupportedPatternError(
+                "Async dict comprehensions are not supported",
+                "Rewrite using an explicit async for loop",
+                line=line,
+                col=col,
+            )
+
+        target_name = node.targets[0].id
+
+        # Initialize accumulator: result = {}
+        init_assign_ast = ast.Assign(
+            targets=[ast.Name(id=target_name, ctx=ast.Store())],
+            value=ast.Dict(keys=[], values=[]),
+            type_comment=None,
+        )
+        ast.copy_location(init_assign_ast, node)
+        ast.fix_missing_locations(init_assign_ast)
+
+        # result[key] = value
+        subscript_target = ast.Subscript(
+            value=ast.Name(id=target_name, ctx=ast.Load()),
+            slice=copy.deepcopy(dictcomp.key),
+            ctx=ast.Store(),
+        )
+        append_assign_ast = ast.Assign(
+            targets=[subscript_target],
+            value=copy.deepcopy(dictcomp.value),
+            type_comment=None,
+        )
+        ast.copy_location(append_assign_ast, node.value)
+        ast.fix_missing_locations(append_assign_ast)
+
+        loop_body: List[ast.stmt] = []
+        if gen.ifs:
+            condition: ast.expr
+            if len(gen.ifs) == 1:
+                condition = copy.deepcopy(gen.ifs[0])
+            else:
+                condition = ast.BoolOp(op=ast.And(), values=[copy.deepcopy(iff) for iff in gen.ifs])
+                ast.copy_location(condition, gen.ifs[0])
+            if_stmt = ast.If(test=condition, body=[append_assign_ast], orelse=[])
+            ast.copy_location(if_stmt, gen.ifs[0])
+            ast.fix_missing_locations(if_stmt)
+            loop_body.append(if_stmt)
+        else:
+            loop_body.append(append_assign_ast)
+
+        loop_ast = ast.For(
+            target=copy.deepcopy(gen.target),
+            iter=copy.deepcopy(gen.iter),
+            body=loop_body,
+            orelse=[],
+            type_comment=None,
+        )
+        ast.copy_location(loop_ast, node)
+        ast.fix_missing_locations(loop_ast)
+
+        statements: List[ir.Statement] = []
+        init_stmt = self._visit_assign(init_assign_ast)
+        if init_stmt:
+            statements.append(init_stmt)
+        statements.extend(self._visit_for(loop_ast))
+
+        return statements
+
+    def _visit_ann_assign(self, node: ast.AnnAssign) -> Optional[ir.Statement]:
+        """Convert annotated assignment to IR when a value is present."""
+        if node.value is None:
+            return None
+
+        assign = ast.Assign(targets=[node.target], value=node.value, type_comment=None)
+        ast.copy_location(assign, node)
+        ast.fix_missing_locations(assign)
+        return self._visit_assign(assign)
+
+    def _visit_assign(self, node: ast.Assign) -> Optional[ir.Statement]:
+        """Convert assignment to IR.
+
+        All assignments with targets use the Assignment statement type.
+        This provides uniform unpacking support for:
+        - Action calls: a, b = @get_pair()
+        - Parallel blocks: a, b = parallel: @x() @y()
+        - Regular expressions: a, b = some_list
+
+        Raises UnsupportedPatternError for:
+        - Constructor calls: x = MyClass(...)
+        - Non-action await: x = await some_func()
+        """
+        stmt = ir.Statement(span=_make_span(node))
+        targets = self._get_assign_targets(node.targets)
+
+        # Check for Pydantic model or dataclass constructor calls
+        # These are converted to dict expressions
+        model_name = self._is_model_constructor(node.value)
+        if model_name and isinstance(node.value, ast.Call):
+            value_expr = self._convert_model_constructor_to_dict(node.value, model_name)
+            assign = ir.Assignment(targets=targets, value=value_expr)
+            stmt.assignment.CopyFrom(assign)
+            return stmt
+
+        # Check for constructor calls in assignment (e.g., x = MyModel(...))
+        # This must come AFTER the model constructor check since models are allowed
+        self._check_constructor_in_assignment(node.value)
+
+        # Check for asyncio.gather() - convert to parallel or spread expression
+        if isinstance(node.value, ast.Await) and isinstance(node.value.value, ast.Call):
+            gather_call = node.value.value
+            if self._is_asyncio_gather_call(gather_call):
+                gather_result = self._convert_asyncio_gather(gather_call)
+                if gather_result is not None:
+                    if isinstance(gather_result, ir.ParallelExpr):
+                        value = ir.Expr(parallel_expr=gather_result, span=_make_span(node))
+                    else:
+                        # SpreadExpr
+                        value = ir.Expr(spread_expr=gather_result, span=_make_span(node))
+                    assign = ir.Assignment(targets=targets, value=value)
+                    stmt.assignment.CopyFrom(assign)
+                    return stmt
+
+        # Check if this is an action call - wrap in Assignment for uniform unpacking
+        action_call = self._extract_action_call(node.value)
+        if action_call:
+            value = ir.Expr(action_call=action_call, span=_make_span(node))
+            assign = ir.Assignment(targets=targets, value=value)
+            stmt.assignment.CopyFrom(assign)
+            return stmt
+
+        # Regular assignment (variables, literals, expressions)
+        value_expr = self._expr_to_ir_with_model_coercion(node.value)
+        if value_expr:
+            assign = ir.Assignment(targets=targets, value=value_expr)
+            stmt.assignment.CopyFrom(assign)
+            return stmt
+
+        return None
+
+    def _visit_expr_stmt(self, node: ast.Expr) -> Optional[ir.Statement]:
+        """Convert expression statement to IR (side effect only, no assignment)."""
+        stmt = ir.Statement(span=_make_span(node))
+
+        # Check for asyncio.gather() - convert to parallel block statement (side effect)
+        if isinstance(node.value, ast.Await) and isinstance(node.value.value, ast.Call):
+            gather_call = node.value.value
+            if self._is_asyncio_gather_call(gather_call):
+                gather_result = self._convert_asyncio_gather(gather_call)
+                if gather_result is not None:
+                    if isinstance(gather_result, ir.ParallelExpr):
+                        # Side effect only - use ParallelBlock statement
+                        parallel = ir.ParallelBlock()
+                        parallel.calls.extend(gather_result.calls)
+                        stmt.parallel_block.CopyFrom(parallel)
+                        return stmt
+                    else:
+                        # SpreadExpr as side effect - wrap in assignment with no targets
+                        # This handles: await asyncio.gather(*[action(x) for x in items])
+                        value = ir.Expr(spread_expr=gather_result, span=_make_span(node))
+                        assign = ir.Assignment(targets=[], value=value)
+                        stmt.assignment.CopyFrom(assign)
+                        return stmt
+
+        # Check if this is an action call (side effect only)
+        action_call = self._extract_action_call(node.value)
+        if action_call:
+            stmt.action_call.CopyFrom(action_call)
+            return stmt
+
+        # Convert list.append(x) to list = list + [x]
+        # This makes the mutation explicit so data flows correctly through the DAG
+        if isinstance(node.value, ast.Call):
+            call = node.value
+            if (
+                isinstance(call.func, ast.Attribute)
+                and call.func.attr == "append"
+                and isinstance(call.func.value, ast.Name)
+                and len(call.args) == 1
+            ):
+                list_name = call.func.value.id
+                append_value = call.args[0]
+                # Create: list = list + [value]
+                list_var = ir.Expr(variable=ir.Variable(name=list_name), span=_make_span(node))
+                value_expr = self._expr_to_ir_with_model_coercion(append_value)
+                if value_expr:
+                    # Create [value] as a list literal
+                    list_literal = ir.Expr(
+                        list=ir.ListExpr(elements=[value_expr]), span=_make_span(node)
+                    )
+                    # Create list + [value]
+                    concat_expr = ir.Expr(
+                        binary_op=ir.BinaryOp(
+                            op=ir.BinaryOperator.BINARY_OP_ADD, left=list_var, right=list_literal
+                        ),
+                        span=_make_span(node),
+                    )
+                    assign = ir.Assignment(targets=[list_name], value=concat_expr)
+                    stmt.assignment.CopyFrom(assign)
+                    return stmt
+
+        # Regular expression
+        expr = self._expr_to_ir_with_model_coercion(node.value)
+        if expr:
+            stmt.expr_stmt.CopyFrom(ir.ExprStmt(expr=expr))
+            return stmt
+
+        return None
+
+    def _visit_for(self, node: ast.For) -> List[ir.Statement]:
+        """Convert for loop to IR.
+
+        The loop body is emitted as a full block so it can contain multiple
+        statements/calls and early `return`.
+        """
+        # Get loop variables
+        loop_vars: List[str] = []
+        if isinstance(node.target, ast.Name):
+            loop_vars.append(node.target.id)
+        elif isinstance(node.target, ast.Tuple):
+            for elt in node.target.elts:
+                if isinstance(elt, ast.Name):
+                    loop_vars.append(elt.id)
+
+        # Get iterable
+        iterable = self._expr_to_ir_with_model_coercion(node.iter)
+        if not iterable:
+            return []
+
+        # Build body statements (recursively transforms nested structures)
+        body_stmts: List[ir.Statement] = []
+        for body_node in node.body:
+            stmts = self._visit_statement(body_node)
+            body_stmts.extend(stmts)
+
+        stmt = ir.Statement(span=_make_span(node))
+        for_loop = ir.ForLoop(
+            loop_vars=loop_vars,
+            iterable=iterable,
+            block_body=ir.Block(statements=body_stmts, span=_make_span(node)),
+        )
+        stmt.for_loop.CopyFrom(for_loop)
+        return [stmt]
+
+    def _visit_while(self, node: ast.While) -> List[ir.Statement]:
+        """Convert while loop to IR.
+
+        The loop body is emitted as a full block so it can contain multiple
+        statements/calls and early `return`.
+        """
+        condition = self._expr_to_ir_with_model_coercion(node.test)
+        if not condition:
+            return []
+
+        body_stmts: List[ir.Statement] = []
+        for body_node in node.body:
+            stmts = self._visit_statement(body_node)
+            body_stmts.extend(stmts)
+
+        stmt = ir.Statement(span=_make_span(node))
+        while_loop = ir.WhileLoop(
+            condition=condition,
+            block_body=ir.Block(statements=body_stmts, span=_make_span(node)),
+        )
+        stmt.while_loop.CopyFrom(while_loop)
+        return [stmt]
+
+    def _detect_accumulator_targets(
+        self, stmts: List[ir.Statement], in_scope_vars: set
+    ) -> List[str]:
+        """Detect out-of-scope variable modifications in for loop body.
+
+        Scans statements for patterns that modify variables defined outside the loop.
+        Returns a list of accumulator variable names that should be set as targets.
+
+        Supported patterns:
+        1. List append: results.append(value) -> "results"
+        2. Dict subscript: result[key] = value -> "result"
+        3. List/set update methods: results.extend(...), results.add(...) -> "results"
+
+        Note: Patterns like `results = results + [x]` and `count = count + 1` create
+        new assignments which are tracked via in_scope_vars and don't need special
+        detection here - they're handled by the regular assignment target logic.
+        """
+        accumulators: List[str] = []
+        seen: set = set()
+
+        for stmt in stmts:
+            var_name = self._extract_accumulator_from_stmt(stmt, in_scope_vars)
+            if var_name and var_name not in seen:
+                accumulators.append(var_name)
+                seen.add(var_name)
+
+            # Check conditionals for accumulator targets in branch bodies
+            if stmt.HasField("conditional"):
+                cond = stmt.conditional
+                branch_blocks: list[ir.Block] = []
+                if cond.HasField("if_branch") and cond.if_branch.HasField("block_body"):
+                    branch_blocks.append(cond.if_branch.block_body)
+                for branch in cond.elif_branches:
+                    if branch.HasField("block_body"):
+                        branch_blocks.append(branch.block_body)
+                if cond.HasField("else_branch") and cond.else_branch.HasField("block_body"):
+                    branch_blocks.append(cond.else_branch.block_body)
+
+                for block in branch_blocks:
+                    for var in self._detect_accumulator_targets(
+                        list(block.statements), in_scope_vars
+                    ):
+                        if var not in seen:
+                            accumulators.append(var)
+                            seen.add(var)
+
+        return accumulators
+
+    def _extract_accumulator_from_stmt(
+        self, stmt: ir.Statement, in_scope_vars: set
+    ) -> Optional[str]:
+        """Extract accumulator variable name from a single statement.
+
+        Returns the variable name if this statement modifies an out-of-scope variable,
+        None otherwise.
+        """
+        # Pattern 1: Method calls like list.append(), dict.update(), set.add()
+        if stmt.HasField("expr_stmt"):
+            expr = stmt.expr_stmt.expr
+            if expr.HasField("function_call"):
+                fn_name = expr.function_call.name
+                # Check for mutating method calls: x.append, x.extend, x.add, x.update, etc.
+                mutating_methods = {
+                    ".append",
+                    ".extend",
+                    ".add",
+                    ".update",
+                    ".insert",
+                    ".pop",
+                    ".remove",
+                    ".clear",
+                }
+                for method in mutating_methods:
+                    if fn_name.endswith(method):
+                        var_name = fn_name[: len(fn_name) - len(method)]
+                        # Only return if it's an out-of-scope variable
+                        if var_name and var_name not in in_scope_vars:
+                            return var_name
+
+        # Pattern 2: Subscript assignment like dict[key] = value
+        if stmt.HasField("assignment"):
+            for target in stmt.assignment.targets:
+                # Check if target is a subscript pattern (contains '[')
+                if "[" in target:
+                    # Extract base variable name (before '[')
+                    var_name = target.split("[")[0]
+                    if var_name and var_name not in in_scope_vars:
+                        return var_name
+
+        # Pattern 3: Self-referential assignment like x = x + [y]
+        # The target variable is used on the RHS, so it must come from outside.
+        # Note: We don't check in_scope_vars here because the assignment itself
+        # would have added the target to in_scope_vars, but it still needs its
+        # previous value from outside the loop body.
+        if stmt.HasField("assignment"):
+            assign = stmt.assignment
+            rhs_vars = self._collect_variables_from_expr(assign.value)
+            for target in assign.targets:
+                if target in rhs_vars:
+                    return target
+
+        return None
+
+    def _collect_variables_from_expr(self, expr: ir.Expr) -> set:
+        """Recursively collect all variable names used in an expression."""
+        vars_found: set = set()
+
+        if expr.HasField("variable"):
+            vars_found.add(expr.variable.name)
+        elif expr.HasField("binary_op"):
+            vars_found.update(self._collect_variables_from_expr(expr.binary_op.left))
+            vars_found.update(self._collect_variables_from_expr(expr.binary_op.right))
+        elif expr.HasField("unary_op"):
+            vars_found.update(self._collect_variables_from_expr(expr.unary_op.operand))
+        elif expr.HasField("list"):
+            for elem in expr.list.elements:
+                vars_found.update(self._collect_variables_from_expr(elem))
+        elif expr.HasField("dict"):
+            for entry in expr.dict.entries:
+                vars_found.update(self._collect_variables_from_expr(entry.key))
+                vars_found.update(self._collect_variables_from_expr(entry.value))
+        elif expr.HasField("index"):
+            vars_found.update(self._collect_variables_from_expr(expr.index.value))
+            vars_found.update(self._collect_variables_from_expr(expr.index.index))
+        elif expr.HasField("dot"):
+            vars_found.update(self._collect_variables_from_expr(expr.dot.object))
+        elif expr.HasField("function_call"):
+            for kwarg in expr.function_call.kwargs:
+                vars_found.update(self._collect_variables_from_expr(kwarg.value))
+        elif expr.HasField("action_call"):
+            for kwarg in expr.action_call.kwargs:
+                vars_found.update(self._collect_variables_from_expr(kwarg.value))
+
+        return vars_found
+
+    def _visit_if(self, node: ast.If) -> List[ir.Statement]:
+        """Convert if statement to IR.
+
+        Normalizes patterns like:
+            if await some_action(...):
+                ...
+        into:
+            __if_cond_n__ = await some_action(...)
+            if __if_cond_n__:
+                ...
+        """
+
+        def normalize_condition(test: ast.expr) -> tuple[List[ir.Statement], Optional[ir.Expr]]:
+            action_call = self._extract_action_call(test)
+            if action_call is None:
+                return ([], self._expr_to_ir_with_model_coercion(test))
+
+            if not isinstance(test, ast.Await):
+                line = getattr(test, "lineno", None)
+                col = getattr(test, "col_offset", None)
+                raise UnsupportedPatternError(
+                    "Action calls inside boolean expressions are not supported in if conditions",
+                    "Assign the awaited action result to a variable, then use the variable in the if condition.",
+                    line=line,
+                    col=col,
+                )
+
+            cond_var = self._ctx.next_implicit_fn_name(prefix="if_cond")
+            assign_stmt = ir.Statement(span=_make_span(test))
+            assign_stmt.assignment.CopyFrom(
+                ir.Assignment(
+                    targets=[cond_var],
+                    value=ir.Expr(action_call=action_call, span=_make_span(test)),
+                )
+            )
+            cond_expr = ir.Expr(variable=ir.Variable(name=cond_var), span=_make_span(test))
+            return ([assign_stmt], cond_expr)
+
+        def visit_body(nodes: list[ast.stmt]) -> List[ir.Statement]:
+            stmts: List[ir.Statement] = []
+            for body_node in nodes:
+                stmts.extend(self._visit_statement(body_node))
+            return stmts
+
+        # Collect if/elif branches as (test_expr, body_nodes)
+        branches: list[tuple[ast.expr, list[ast.stmt], ast.AST]] = [(node.test, node.body, node)]
+        current = node
+        while current.orelse and len(current.orelse) == 1 and isinstance(current.orelse[0], ast.If):
+            elif_node = current.orelse[0]
+            branches.append((elif_node.test, elif_node.body, elif_node))
+            current = elif_node
+
+        else_nodes = current.orelse
+
+        normalized: list[
+            tuple[List[ir.Statement], Optional[ir.Expr], List[ir.Statement], ast.AST]
+        ] = []
+        for test_expr, body_nodes, span_node in branches:
+            prefix, cond = normalize_condition(test_expr)
+            normalized.append((prefix, cond, visit_body(body_nodes), span_node))
+
+        else_body = visit_body(else_nodes) if else_nodes else []
+
+        # If any non-first branch needs normalization, preserve Python semantics by nesting.
+        requires_nested = any(prefix for prefix, _, _, _ in normalized[1:])
+
+        def build_conditional_stmt(
+            condition: ir.Expr,
+            then_body: List[ir.Statement],
+            else_body_statements: List[ir.Statement],
+            span_node: ast.AST,
+        ) -> ir.Statement:
+            conditional_stmt = ir.Statement(span=_make_span(span_node))
+            if_branch = ir.IfBranch(
+                condition=condition,
+                block_body=ir.Block(statements=then_body, span=_make_span(span_node)),
+                span=_make_span(span_node),
+            )
+            conditional = ir.Conditional(if_branch=if_branch)
+            if else_body_statements:
+                else_branch = ir.ElseBranch(
+                    block_body=ir.Block(
+                        statements=else_body_statements,
+                        span=_make_span(span_node),
+                    ),
+                    span=_make_span(span_node),
+                )
+                conditional.else_branch.CopyFrom(else_branch)
+            conditional_stmt.conditional.CopyFrom(conditional)
+            return conditional_stmt
+
+        if requires_nested:
+            nested_else: List[ir.Statement] = else_body
+            for prefix, cond, then_body, span_node in reversed(normalized):
+                if cond is None:
+                    continue
+                nested_if_stmt = build_conditional_stmt(
+                    condition=cond,
+                    then_body=then_body,
+                    else_body_statements=nested_else,
+                    span_node=span_node,
+                )
+                nested_else = [*prefix, nested_if_stmt]
+            return nested_else
+
+        # Flat conditional with elif/else (original behavior), plus optional prefix for the if guard.
+        if_prefix, if_condition, if_body, if_span_node = normalized[0]
+        if if_condition is None:
+            return []
+
+        conditional_stmt = ir.Statement(span=_make_span(if_span_node))
+        if_branch = ir.IfBranch(
+            condition=if_condition,
+            block_body=ir.Block(statements=if_body, span=_make_span(if_span_node)),
+            span=_make_span(if_span_node),
+        )
+        conditional = ir.Conditional(if_branch=if_branch)
+
+        for _, elif_condition, elif_body, elif_span_node in normalized[1:]:
+            if elif_condition is None:
+                continue
+            elif_branch = ir.ElifBranch(
+                condition=elif_condition,
+                block_body=ir.Block(statements=elif_body, span=_make_span(elif_span_node)),
+                span=_make_span(elif_span_node),
+            )
+            conditional.elif_branches.append(elif_branch)
+
+        if else_body:
+            else_branch = ir.ElseBranch(
+                block_body=ir.Block(statements=else_body, span=_make_span(if_span_node)),
+                span=_make_span(if_span_node),
+            )
+            conditional.else_branch.CopyFrom(else_branch)
+
+        conditional_stmt.conditional.CopyFrom(conditional)
+        return [*if_prefix, conditional_stmt]
+
+    def _collect_assigned_vars(self, stmts: List[ir.Statement]) -> set:
+        """Collect all variable names assigned in a list of statements."""
+        assigned = set()
+        for stmt in stmts:
+            if stmt.HasField("assignment"):
+                assigned.update(stmt.assignment.targets)
+        return assigned
+
+    def _collect_assigned_vars_in_order(self, stmts: List[ir.Statement]) -> list[str]:
+        """Collect assigned variable names in statement order (deduplicated)."""
+        assigned: list[str] = []
+        seen: set[str] = set()
+
+        for stmt in stmts:
+            if stmt.HasField("assignment"):
+                for target in stmt.assignment.targets:
+                    if target not in seen:
+                        seen.add(target)
+                        assigned.append(target)
+
+            if stmt.HasField("conditional"):
+                cond = stmt.conditional
+                if cond.HasField("if_branch") and cond.if_branch.HasField("block_body"):
+                    for target in self._collect_assigned_vars_in_order(
+                        list(cond.if_branch.block_body.statements)
+                    ):
+                        if target not in seen:
+                            seen.add(target)
+                            assigned.append(target)
+                for elif_branch in cond.elif_branches:
+                    if elif_branch.HasField("block_body"):
+                        for target in self._collect_assigned_vars_in_order(
+                            list(elif_branch.block_body.statements)
+                        ):
+                            if target not in seen:
+                                seen.add(target)
+                                assigned.append(target)
+                if cond.HasField("else_branch") and cond.else_branch.HasField("block_body"):
+                    for target in self._collect_assigned_vars_in_order(
+                        list(cond.else_branch.block_body.statements)
+                    ):
+                        if target not in seen:
+                            seen.add(target)
+                            assigned.append(target)
+
+            if stmt.HasField("for_loop") and stmt.for_loop.HasField("block_body"):
+                for target in self._collect_assigned_vars_in_order(
+                    list(stmt.for_loop.block_body.statements)
+                ):
+                    if target not in seen:
+                        seen.add(target)
+                        assigned.append(target)
+
+            if stmt.HasField("while_loop") and stmt.while_loop.HasField("block_body"):
+                for target in self._collect_assigned_vars_in_order(
+                    list(stmt.while_loop.block_body.statements)
+                ):
+                    if target not in seen:
+                        seen.add(target)
+                        assigned.append(target)
+
+            if stmt.HasField("try_except"):
+                try_block = stmt.try_except.try_block
+                if try_block.HasField("span"):
+                    for target in self._collect_assigned_vars_in_order(list(try_block.statements)):
+                        if target not in seen:
+                            seen.add(target)
+                            assigned.append(target)
+                for handler in stmt.try_except.handlers:
+                    if handler.HasField("block_body"):
+                        for target in self._collect_assigned_vars_in_order(
+                            list(handler.block_body.statements)
+                        ):
+                            if target not in seen:
+                                seen.add(target)
+                                assigned.append(target)
+
+        return assigned
+
+    def _collect_variables_from_block(self, block: ir.Block) -> list[str]:
+        return self._collect_variables_from_statements(list(block.statements))
+
+    def _collect_variables_from_statements(self, stmts: List[ir.Statement]) -> list[str]:
+        """Collect variable references from statements in encounter order."""
+        vars_found: list[str] = []
+        seen: set[str] = set()
+
+        for stmt in stmts:
+            if stmt.HasField("assignment") and stmt.assignment.HasField("value"):
+                for var in self._collect_variables_from_expr(stmt.assignment.value):
+                    if var not in seen:
+                        seen.add(var)
+                        vars_found.append(var)
+
+            if stmt.HasField("return_stmt") and stmt.return_stmt.HasField("value"):
+                for var in self._collect_variables_from_expr(stmt.return_stmt.value):
+                    if var not in seen:
+                        seen.add(var)
+                        vars_found.append(var)
+
+            if stmt.HasField("action_call"):
+                expr = ir.Expr(action_call=stmt.action_call, span=stmt.span)
+                for var in self._collect_variables_from_expr(expr):
+                    if var not in seen:
+                        seen.add(var)
+                        vars_found.append(var)
+
+            if stmt.HasField("expr_stmt"):
+                for var in self._collect_variables_from_expr(stmt.expr_stmt.expr):
+                    if var not in seen:
+                        seen.add(var)
+                        vars_found.append(var)
+
+            if stmt.HasField("conditional"):
+                cond = stmt.conditional
+                if cond.HasField("if_branch"):
+                    if cond.if_branch.HasField("condition"):
+                        for var in self._collect_variables_from_expr(cond.if_branch.condition):
+                            if var not in seen:
+                                seen.add(var)
+                                vars_found.append(var)
+                    if cond.if_branch.HasField("block_body"):
+                        for var in self._collect_variables_from_block(cond.if_branch.block_body):
+                            if var not in seen:
+                                seen.add(var)
+                                vars_found.append(var)
+                for elif_branch in cond.elif_branches:
+                    if elif_branch.HasField("condition"):
+                        for var in self._collect_variables_from_expr(elif_branch.condition):
+                            if var not in seen:
+                                seen.add(var)
+                                vars_found.append(var)
+                    if elif_branch.HasField("block_body"):
+                        for var in self._collect_variables_from_block(elif_branch.block_body):
+                            if var not in seen:
+                                seen.add(var)
+                                vars_found.append(var)
+                if cond.HasField("else_branch") and cond.else_branch.HasField("block_body"):
+                    for var in self._collect_variables_from_block(cond.else_branch.block_body):
+                        if var not in seen:
+                            seen.add(var)
+                            vars_found.append(var)
+
+            if stmt.HasField("for_loop"):
+                fl = stmt.for_loop
+                if fl.HasField("iterable"):
+                    for var in self._collect_variables_from_expr(fl.iterable):
+                        if var not in seen:
+                            seen.add(var)
+                            vars_found.append(var)
+                if fl.HasField("block_body"):
+                    for var in self._collect_variables_from_block(fl.block_body):
+                        if var not in seen:
+                            seen.add(var)
+                            vars_found.append(var)
+
+            if stmt.HasField("while_loop"):
+                wl = stmt.while_loop
+                if wl.HasField("condition"):
+                    for var in self._collect_variables_from_expr(wl.condition):
+                        if var not in seen:
+                            seen.add(var)
+                            vars_found.append(var)
+                if wl.HasField("block_body"):
+                    for var in self._collect_variables_from_block(wl.block_body):
+                        if var not in seen:
+                            seen.add(var)
+                            vars_found.append(var)
+
+            if stmt.HasField("try_except"):
+                te = stmt.try_except
+                if te.HasField("try_block"):
+                    for var in self._collect_variables_from_block(te.try_block):
+                        if var not in seen:
+                            seen.add(var)
+                            vars_found.append(var)
+                for handler in te.handlers:
+                    if handler.HasField("block_body"):
+                        for var in self._collect_variables_from_block(handler.block_body):
+                            if var not in seen:
+                                seen.add(var)
+                                vars_found.append(var)
+
+            if stmt.HasField("parallel_block"):
+                for call in stmt.parallel_block.calls:
+                    if call.HasField("action"):
+                        for kwarg in call.action.kwargs:
+                            for var in self._collect_variables_from_expr(kwarg.value):
+                                if var not in seen:
+                                    seen.add(var)
+                                    vars_found.append(var)
+                    elif call.HasField("function"):
+                        for kwarg in call.function.kwargs:
+                            for var in self._collect_variables_from_expr(kwarg.value):
+                                if var not in seen:
+                                    seen.add(var)
+                                    vars_found.append(var)
+
+            if stmt.HasField("spread_action"):
+                spread = stmt.spread_action
+                if spread.HasField("collection"):
+                    for var in self._collect_variables_from_expr(spread.collection):
+                        if var not in seen:
+                            seen.add(var)
+                            vars_found.append(var)
+                if spread.HasField("action"):
+                    for kwarg in spread.action.kwargs:
+                        for var in self._collect_variables_from_expr(kwarg.value):
+                            if var not in seen:
+                                seen.add(var)
+                                vars_found.append(var)
+
+        return vars_found
+
+    def _visit_try(self, node: ast.Try) -> List[ir.Statement]:
+        """Convert try/except to IR with full block bodies."""
+        # Build try body statements (recursively transforms nested structures)
+        try_body: List[ir.Statement] = []
+        for body_node in node.body:
+            stmts = self._visit_statement(body_node)
+            try_body.extend(stmts)
+
+        # Build exception handlers (with wrapping if needed)
+        handlers: List[ir.ExceptHandler] = []
+        for handler in node.handlers:
+            exception_types: List[str] = []
+            if handler.type:
+                if isinstance(handler.type, ast.Name):
+                    exception_types.append(handler.type.id)
+                elif isinstance(handler.type, ast.Tuple):
+                    for elt in handler.type.elts:
+                        if isinstance(elt, ast.Name):
+                            exception_types.append(elt.id)
+
+            # Build handler body (recursively transforms nested structures)
+            handler_body: List[ir.Statement] = []
+            for handler_node in handler.body:
+                stmts = self._visit_statement(handler_node)
+                handler_body.extend(stmts)
+
+            except_handler = ir.ExceptHandler(
+                exception_types=exception_types,
+                block_body=ir.Block(statements=handler_body, span=_make_span(handler)),
+                span=_make_span(handler),
+            )
+            if handler.name:
+                except_handler.exception_var = handler.name
+            handlers.append(except_handler)
+
+        # Build the try/except statement
+        try_stmt = ir.Statement(span=_make_span(node))
+        try_except = ir.TryExcept(
+            handlers=handlers,
+            try_block=ir.Block(statements=try_body, span=_make_span(node)),
+        )
+        try_stmt.try_except.CopyFrom(try_except)
+
+        return [try_stmt]
+
+    def _count_calls(self, stmts: List[ir.Statement]) -> int:
+        """Count action calls and function calls in statements.
+
+        Both action calls and function calls (including synthetic functions)
+        count toward the limit of one call per control flow body.
+        """
+        count = 0
+        for stmt in stmts:
+            if stmt.HasField("action_call"):
+                count += 1
+            elif stmt.HasField("assignment"):
+                # Check if assignment value is an action call or function call
+                if stmt.assignment.value.HasField("action_call"):
+                    count += 1
+                elif stmt.assignment.value.HasField("function_call"):
+                    count += 1
+            elif stmt.HasField("expr_stmt"):
+                # Check if expression is a function call
+                if stmt.expr_stmt.expr.HasField("function_call"):
+                    count += 1
+        return count
+
+    def _wrap_body_as_function(
+        self,
+        body: List[ir.Statement],
+        prefix: str,
+        node: ast.AST,
+        inputs: Optional[List[str]] = None,
+        modified_vars: Optional[List[str]] = None,
+    ) -> List[ir.Statement]:
+        """Wrap a body with multiple calls into a synthetic function.
+
+        Args:
+            body: The statements to wrap
+            prefix: Name prefix for the synthetic function
+            node: AST node for span information
+            inputs: Variables to pass as inputs (e.g., loop variables)
+            modified_vars: Out-of-scope variables modified in the body.
+                          These are added as inputs AND returned as outputs,
+                          enabling functional transformation of external state.
+
+        Returns a list containing a single function call statement (or assignment
+        if modified_vars are present).
+        """
+        fn_name = self._ctx.next_implicit_fn_name(prefix)
+        fn_inputs = list(inputs or [])
+
+        # Add modified variables as inputs (they need to be passed in)
+        modified_vars = modified_vars or []
+        for var in modified_vars:
+            if var not in fn_inputs:
+                fn_inputs.append(var)
+
+        # If there are modified variables, add a return statement for them
+        wrapped_body = list(body)
+        if modified_vars:
+            # Create return statement: return (var1, var2, ...) or return var1
+            if len(modified_vars) == 1:
+                return_expr = ir.Expr(
+                    variable=ir.Variable(name=modified_vars[0]),
+                    span=_make_span(node),
+                )
+            else:
+                # Return as list (tuples are represented as lists in IR)
+                return_expr = ir.Expr(
+                    list=ir.ListExpr(
+                        elements=[ir.Expr(variable=ir.Variable(name=var)) for var in modified_vars]
+                    ),
+                    span=_make_span(node),
+                )
+            return_stmt = ir.Statement(span=_make_span(node))
+            return_stmt.return_stmt.CopyFrom(ir.ReturnStmt(value=return_expr))
+            wrapped_body.append(return_stmt)
+
+        # Create the synthetic function
+        implicit_fn = ir.FunctionDef(
+            name=fn_name,
+            io=ir.IoDecl(inputs=fn_inputs, outputs=modified_vars),
+            body=ir.Block(statements=wrapped_body),
+            span=_make_span(node),
+        )
+        self._ctx.implicit_functions.append(implicit_fn)
+
+        # Create a function call expression
+        kwargs = [
+            ir.Kwarg(name=var, value=ir.Expr(variable=ir.Variable(name=var))) for var in fn_inputs
+        ]
+        fn_call_expr = ir.Expr(
+            function_call=ir.FunctionCall(name=fn_name, kwargs=kwargs),
+            span=_make_span(node),
+        )
+
+        # If there are modified variables, create an assignment statement
+        # so the returned values are assigned back to the variables
+        call_stmt = ir.Statement(span=_make_span(node))
+        if modified_vars:
+            # Create assignment: var1, var2 = fn(...) or var1 = fn(...)
+            assign = ir.Assignment(value=fn_call_expr)
+            assign.targets.extend(modified_vars)
+            call_stmt.assignment.CopyFrom(assign)
+        else:
+            call_stmt.expr_stmt.CopyFrom(ir.ExprStmt(expr=fn_call_expr))
+
+        return [call_stmt]
+
+    def _visit_return(self, node: ast.Return) -> List[ir.Statement]:
+        """Convert return statement to IR.
+
+        Return statements should only contain variables or literals, not action calls.
+        If the return contains an action call, we normalize it:
+            return await action()
+        becomes:
+            _return_tmp = await action()
+            return _return_tmp
+
+        Constructor calls (like return MyModel(...)) are not supported and will
+        raise an error with a recommendation to use an @action instead.
+        """
+        if node.value:
+            # Check for constructor calls in return (e.g., return MyModel(...))
+            self._check_constructor_in_return(node.value)
+
+            # Check if returning an action call - normalize to assignment + return
+            action_call = self._extract_action_call(node.value)
+            if action_call:
+                # Create a temporary variable for the action result
+                tmp_var = "_return_tmp"
+
+                # Create assignment: _return_tmp = await action()
+                assign_stmt = ir.Statement(span=_make_span(node))
+                value = ir.Expr(action_call=action_call, span=_make_span(node))
+                assign = ir.Assignment(targets=[tmp_var], value=value)
+                assign_stmt.assignment.CopyFrom(assign)
+
+                # Create return: return _return_tmp
+                return_stmt = ir.Statement(span=_make_span(node))
+                var_expr = ir.Expr(variable=ir.Variable(name=tmp_var), span=_make_span(node))
+                ret = ir.ReturnStmt(value=var_expr)
+                return_stmt.return_stmt.CopyFrom(ret)
+
+                return [assign_stmt, return_stmt]
+
+            # Normalize return of function calls into assignment + return
+            expr = self._expr_to_ir_with_model_coercion(node.value)
+            if expr and expr.HasField("function_call"):
+                tmp_var = self._ctx.next_implicit_fn_name(prefix="return_tmp")
+
+                assign_stmt = ir.Statement(span=_make_span(node))
+                assign_stmt.assignment.CopyFrom(ir.Assignment(targets=[tmp_var], value=expr))
+
+                return_stmt = ir.Statement(span=_make_span(node))
+                var_expr = ir.Expr(variable=ir.Variable(name=tmp_var), span=_make_span(node))
+                return_stmt.return_stmt.CopyFrom(ir.ReturnStmt(value=var_expr))
+                return [assign_stmt, return_stmt]
+
+            # Regular return with expression (variable, literal, etc.)
+            if expr:
+                stmt = ir.Statement(span=_make_span(node))
+                return_stmt = ir.ReturnStmt(value=expr)
+                stmt.return_stmt.CopyFrom(return_stmt)
+                return [stmt]
+
+        # Return with no value
+        stmt = ir.Statement(span=_make_span(node))
+        stmt.return_stmt.CopyFrom(ir.ReturnStmt())
+        return [stmt]
+
+    def _visit_break(self, node: ast.Break) -> List[ir.Statement]:
+        """Convert break statement to IR."""
+        stmt = ir.Statement(span=_make_span(node))
+        stmt.break_stmt.CopyFrom(ir.BreakStmt())
+        return [stmt]
+
+    def _visit_continue(self, node: ast.Continue) -> List[ir.Statement]:
+        """Convert continue statement to IR."""
+        stmt = ir.Statement(span=_make_span(node))
+        stmt.continue_stmt.CopyFrom(ir.ContinueStmt())
+        return [stmt]
+
+    def _visit_aug_assign(self, node: ast.AugAssign) -> List[ir.Statement]:
+        """Convert augmented assignment (+=, -=, etc.) to IR."""
+        # For now, we can represent this as a regular assignment with binary op
+        # target op= value  ->  target = target op value
+        stmt = ir.Statement(span=_make_span(node))
+
+        targets: List[str] = []
+        if isinstance(node.target, ast.Name):
+            targets.append(node.target.id)
+
+        left = self._expr_to_ir_with_model_coercion(node.target)
+        right = self._expr_to_ir_with_model_coercion(node.value)
+        if right and right.HasField("function_call"):
+            tmp_var = self._ctx.next_implicit_fn_name(prefix="aug_tmp")
+
+            assign_tmp = ir.Statement(span=_make_span(node))
+            assign_tmp.assignment.CopyFrom(
+                ir.Assignment(
+                    targets=[tmp_var],
+                    value=ir.Expr(function_call=right.function_call, span=_make_span(node)),
+                )
+            )
+
+            if left:
+                op = _bin_op_to_ir(node.op)
+                if op:
+                    binary = ir.BinaryOp(
+                        left=left,
+                        op=op,
+                        right=ir.Expr(variable=ir.Variable(name=tmp_var)),
+                    )
+                    value = ir.Expr(binary_op=binary)
+                    assign = ir.Assignment(targets=targets, value=value)
+                    stmt.assignment.CopyFrom(assign)
+                    return [assign_tmp, stmt]
+            return [assign_tmp]
+
+        if left and right:
+            op = _bin_op_to_ir(node.op)
+            if op:
+                binary = ir.BinaryOp(left=left, op=op, right=right)
+                value = ir.Expr(binary_op=binary)
+                assign = ir.Assignment(targets=targets, value=value)
+                stmt.assignment.CopyFrom(assign)
+                return [stmt]
+
+        return []
+
+    def _collect_function_inputs(self, node: ast.FunctionDef | ast.AsyncFunctionDef) -> List[str]:
+        """Collect workflow inputs from function parameters, including kw-only args."""
+        args: List[str] = []
+        seen: set[str] = set()
+
+        ordered_args = list(node.args.posonlyargs) + list(node.args.args)
+        if ordered_args and ordered_args[0].arg == "self":
+            ordered_args = ordered_args[1:]
+
+        for arg in ordered_args:
+            if arg.arg not in seen:
+                args.append(arg.arg)
+                seen.add(arg.arg)
+
+        if node.args.vararg and node.args.vararg.arg not in seen:
+            args.append(node.args.vararg.arg)
+            seen.add(node.args.vararg.arg)
+
+        for arg in node.args.kwonlyargs:
+            if arg.arg not in seen:
+                args.append(arg.arg)
+                seen.add(arg.arg)
+
+        if node.args.kwarg and node.args.kwarg.arg not in seen:
+            args.append(node.args.kwarg.arg)
+            seen.add(node.args.kwarg.arg)
+
+        return args
+
+    def _check_constructor_in_return(self, node: ast.expr) -> None:
+        """Check for constructor calls in return statements.
+
+        Raises UnsupportedPatternError if the return value is a class instantiation
+        like: return MyModel(field=value)
+
+        This is not supported because the workflow IR cannot serialize arbitrary
+        object instantiation. Users should use an @action to create objects.
+        """
+        # Skip if it's an await (action call) - those are fine
+        if isinstance(node, ast.Await):
+            return
+
+        # Check for direct Call that looks like a constructor
+        if isinstance(node, ast.Call):
+            func_name = self._get_constructor_name(node.func)
+            if func_name and self._looks_like_constructor(func_name, node):
+                line = getattr(node, "lineno", None)
+                col = getattr(node, "col_offset", None)
+                raise UnsupportedPatternError(
+                    f"Returning constructor call '{func_name}(...)' is not supported",
+                    RECOMMENDATIONS["constructor_return"],
+                    line=line,
+                    col=col,
+                )
+
+    def _check_constructor_in_assignment(self, node: ast.expr) -> None:
+        """Check for constructor calls in assignments.
+
+        Raises UnsupportedPatternError if the assignment value is a class instantiation
+        like: result = MyModel(field=value)
+
+        This is not supported because the workflow IR cannot serialize arbitrary
+        object instantiation. Users should use an @action to create objects.
+        """
+        # Skip if it's an await (action call) - those are fine
+        if isinstance(node, ast.Await):
+            return
+
+        # Check for direct Call that looks like a constructor
+        if isinstance(node, ast.Call):
+            func_name = self._get_constructor_name(node.func)
+            if func_name and self._looks_like_constructor(func_name, node):
+                line = getattr(node, "lineno", None)
+                col = getattr(node, "col_offset", None)
+                raise UnsupportedPatternError(
+                    f"Assigning constructor call '{func_name}(...)' is not supported",
+                    RECOMMENDATIONS["constructor_assignment"],
+                    line=line,
+                    col=col,
+                )
+
+    def _get_constructor_name(self, func: ast.expr) -> Optional[str]:
+        """Get the name from a function expression if it looks like a constructor."""
+        if isinstance(func, ast.Name):
+            return func.id
+        elif isinstance(func, ast.Attribute):
+            return func.attr
+        return None
+
+    def _looks_like_constructor(self, func_name: str, call: ast.Call) -> bool:
+        """Check if a function call looks like a class constructor.
+
+        A constructor is identified by:
+        1. Name starts with uppercase (PEP8 convention for classes)
+        2. It's not a known action
+        3. It's not a known builtin like String operations
+        4. It's not a known Pydantic model or dataclass (those are allowed)
+
+        This is a heuristic - we can't perfectly distinguish constructors
+        from functions without full type information.
+        """
+        # Check if first letter is uppercase (class naming convention)
+        if not func_name or not func_name[0].isupper():
+            return False
+
+        # If it's a known action, it's not a constructor
+        if func_name in self._action_defs:
+            return False
+
+        # If it's a known Pydantic model or dataclass, allow it
+        # (it will be converted to a dict expression)
+        if func_name in self._model_defs:
+            return False
+
+        # Common builtins that start with uppercase but aren't constructors
+        # (these are rarely used in workflow code but let's be safe)
+        builtin_exceptions = {"True", "False", "None", "Ellipsis"}
+        if func_name in builtin_exceptions:
+            return False
+
+        return True
+
+    def _is_model_constructor(self, node: ast.expr) -> Optional[str]:
+        """Check if an expression is a Pydantic model or dataclass constructor call.
+
+        Returns the model name if it is, None otherwise.
+        """
+        if not isinstance(node, ast.Call):
+            return None
+
+        func_name = self._get_constructor_name(node.func)
+        if func_name and func_name in self._model_defs:
+            return func_name
+
+        return None
+
+    def _convert_model_constructor_to_dict(self, node: ast.Call, model_name: str) -> ir.Expr:
+        """Convert a Pydantic model or dataclass constructor call to a dict expression.
+
+        For example:
+            MyModel(field1=value1, field2=value2)
+        becomes:
+            {"field1": value1, "field2": value2}
+
+        Default values from the model definition are included for fields not
+        explicitly provided in the constructor call.
+        """
+        model_def = self._model_defs[model_name]
+        entries: List[ir.DictEntry] = []
+
+        # Track which fields were explicitly provided
+        provided_fields: Set[str] = set()
+
+        # First, add all explicitly provided kwargs
+        for kw in node.keywords:
+            if kw.arg is None:
+                # **kwargs expansion - not supported
+                line = getattr(node, "lineno", None)
+                col = getattr(node, "col_offset", None)
+                raise UnsupportedPatternError(
+                    f"Model constructor '{model_name}' with **kwargs is not supported",
+                    "Use explicit keyword arguments instead of **kwargs.",
+                    line=line,
+                    col=col,
+                )
+
+            provided_fields.add(kw.arg)
+            key_expr = ir.Expr()
+            key_literal = ir.Literal()
+            key_literal.string_value = kw.arg
+            key_expr.literal.CopyFrom(key_literal)
+
+            value_expr = self._expr_to_ir_with_model_coercion(kw.value)
+            if value_expr is None:
+                # If we can't convert the value, we need to raise an error
+                line = getattr(node, "lineno", None)
+                col = getattr(node, "col_offset", None)
+                raise UnsupportedPatternError(
+                    f"Cannot convert value for field '{kw.arg}' in '{model_name}'",
+                    "Use simpler expressions (literals, variables, dicts, lists).",
+                    line=line,
+                    col=col,
+                )
+
+            entries.append(ir.DictEntry(key=key_expr, value=value_expr))
+
+        # Handle positional arguments (dataclasses support this)
+        if node.args:
+            # For dataclasses, positional args map to fields in order
+            field_names = list(model_def.fields.keys())
+            for i, arg in enumerate(node.args):
+                if i >= len(field_names):
+                    line = getattr(node, "lineno", None)
+                    col = getattr(node, "col_offset", None)
+                    raise UnsupportedPatternError(
+                        f"Too many positional arguments for '{model_name}'",
+                        "Use keyword arguments for clarity.",
+                        line=line,
+                        col=col,
+                    )
+
+                field_name = field_names[i]
+                provided_fields.add(field_name)
+
+                key_expr = ir.Expr()
+                key_literal = ir.Literal()
+                key_literal.string_value = field_name
+                key_expr.literal.CopyFrom(key_literal)
+
+                value_expr = self._expr_to_ir_with_model_coercion(arg)
+                if value_expr is None:
+                    line = getattr(node, "lineno", None)
+                    col = getattr(node, "col_offset", None)
+                    raise UnsupportedPatternError(
+                        f"Cannot convert positional argument for field '{field_name}' in '{model_name}'",
+                        "Use simpler expressions (literals, variables, dicts, lists).",
+                        line=line,
+                        col=col,
+                    )
+
+                entries.append(ir.DictEntry(key=key_expr, value=value_expr))
+
+        # Add default values for fields not explicitly provided
+        for field_name, field_def in model_def.fields.items():
+            if field_name in provided_fields:
+                continue
+
+            if field_def.has_default:
+                key_expr = ir.Expr()
+                key_literal = ir.Literal()
+                key_literal.string_value = field_name
+                key_expr.literal.CopyFrom(key_literal)
+
+                # Convert the default value to an IR literal
+                default_literal = _constant_to_literal(field_def.default_value)
+                if default_literal is None:
+                    # Can't serialize this default - skip it
+                    # (it's probably a complex object like a list factory)
+                    continue
+
+                value_expr = ir.Expr()
+                value_expr.literal.CopyFrom(default_literal)
+
+                entries.append(ir.DictEntry(key=key_expr, value=value_expr))
+
+        result = ir.Expr(span=_make_span(node))
+        result.dict.CopyFrom(ir.DictExpr(entries=entries))
+        return result
+
+    def _check_non_action_await(self, node: ast.Await) -> None:
+        """Check if an await is for a non-action function.
+
+        Note: We can only reliably detect non-action awaits for functions defined
+        in the same module. Actions imported from other modules will pass through
+        and may fail at runtime if they're not actually actions.
+
+        For now, we only check against common builtins and known non-action patterns.
+        A runtime check will catch functions that aren't registered actions.
+        """
+        awaited = node.value
+        if not isinstance(awaited, ast.Call):
+            return
+
+        # Skip special cases that are handled elsewhere
+        if self._is_run_action_call(awaited):
+            return
+        if self._is_asyncio_sleep_call(awaited):
+            return
+        if self._is_asyncio_gather_call(awaited):
+            return
+
+        # Get the function name
+        func_name = None
+        if isinstance(awaited.func, ast.Name):
+            func_name = awaited.func.id
+        elif isinstance(awaited.func, ast.Attribute):
+            func_name = awaited.func.attr
+
+        if not func_name:
+            return
+
+        # Only raise error for functions defined in THIS module that we know
+        # are NOT actions (i.e., async functions without @action decorator)
+        # We can't reliably detect imported non-actions without full type info.
+        #
+        # The check works by looking at _module_functions which contains
+        # functions defined in the same module as the workflow.
+        if func_name in getattr(self, "_module_functions", set()):
+            if func_name not in self._action_defs:
+                line = getattr(node, "lineno", None)
+                col = getattr(node, "col_offset", None)
+                raise UnsupportedPatternError(
+                    f"Awaiting non-action function '{func_name}()' is not supported",
+                    RECOMMENDATIONS["non_action_call"],
+                    line=line,
+                    col=col,
+                )
+
+    def _check_sync_function_call(self, node: ast.Call) -> None:
+        """Check for synchronous function calls that should be in actions.
+
+        Common patterns like len(), str(), etc. are not supported in workflow code.
+        """
+        func_name = None
+        if isinstance(node.func, ast.Name):
+            func_name = node.func.id
+        elif isinstance(node.func, ast.Attribute):
+            # Method calls on objects - check the method name
+            func_name = node.func.attr
+
+        if not func_name:
+            return
+
+        # Builtins that users commonly try to use
+        common_builtins = {
+            "len",
+            "str",
+            "int",
+            "float",
+            "bool",
+            "list",
+            "dict",
+            "set",
+            "tuple",
+            "sum",
+            "min",
+            "max",
+            "sorted",
+            "reversed",
+            "enumerate",
+            "zip",
+            "map",
+            "filter",
+            "range",
+            "abs",
+            "round",
+            "print",
+            "type",
+            "isinstance",
+            "hasattr",
+            "getattr",
+            "setattr",
+            "open",
+            "format",
+        }
+
+        if func_name in common_builtins:
+            line = getattr(node, "lineno", None)
+            col = getattr(node, "col_offset", None)
+            raise UnsupportedPatternError(
+                f"Calling built-in function '{func_name}()' directly is not supported",
+                RECOMMENDATIONS["builtin_call"],
+                line=line,
+                col=col,
+            )
+
+    def _extract_action_call(self, node: ast.expr) -> Optional[ir.ActionCall]:
+        """Extract an action call from an expression if present.
+
+        Also validates that awaited calls are actually @action decorated functions.
+        Raises UnsupportedPatternError if awaiting a non-action function.
+        """
+        if not isinstance(node, ast.Await):
+            return None
+
+        awaited = node.value
+        # Handle self.run_action(...) wrapper
+        if isinstance(awaited, ast.Call):
+            if self._is_run_action_call(awaited):
+                # Extract the actual action call from run_action
+                if awaited.args:
+                    action_call = self._extract_action_call_from_awaitable(awaited.args[0])
+                    if action_call:
+                        # Extract policies from run_action kwargs (retry, timeout)
+                        self._extract_policies_from_run_action(awaited, action_call)
+                    return action_call
+            # Check for asyncio.sleep() - convert to @sleep action
+            if self._is_asyncio_sleep_call(awaited):
+                return self._convert_asyncio_sleep_to_action(awaited)
+            # Try to extract as action call
+            action_call = self._extract_action_call_from_call(awaited)
+            if action_call:
+                return action_call
+
+            # If we get here, it's an await of a non-action function
+            self._check_non_action_await(node)
+            return None
+
+        return None
+
+    def _is_run_action_call(self, node: ast.Call) -> bool:
+        """Check if this is a self.run_action(...) call."""
+        if isinstance(node.func, ast.Attribute):
+            return node.func.attr == "run_action"
+        return False
+
+    def _extract_policies_from_run_action(
+        self, run_action_call: ast.Call, action_call: ir.ActionCall
+    ) -> None:
+        """Extract retry and timeout policies from run_action kwargs.
+
+        Parses patterns like:
+        - self.run_action(action(), retry=RetryPolicy(attempts=3))
+        - self.run_action(action(), timeout=timedelta(seconds=30))
+        - self.run_action(action(), timeout=60)
+        """
+        for kw in run_action_call.keywords:
+            if kw.arg == "retry":
+                retry_policy = self._parse_retry_policy(kw.value)
+                if retry_policy:
+                    policy_bracket = ir.PolicyBracket()
+                    policy_bracket.retry.CopyFrom(retry_policy)
+                    action_call.policies.append(policy_bracket)
+            elif kw.arg == "timeout":
+                timeout_policy = self._parse_timeout_policy(kw.value)
+                if timeout_policy:
+                    policy_bracket = ir.PolicyBracket()
+                    policy_bracket.timeout.CopyFrom(timeout_policy)
+                    action_call.policies.append(policy_bracket)
+
+    def _parse_retry_policy(self, node: ast.expr) -> Optional[ir.RetryPolicy]:
+        """Parse a RetryPolicy(...) call into IR.
+
+        Supports:
+        - RetryPolicy(attempts=3)
+        - RetryPolicy(attempts=3, exception_types=["ValueError"])
+        - RetryPolicy(attempts=3, backoff_seconds=5)
+        - self.retry_policy (instance attribute reference)
+        """
+        # Handle self.attr pattern - look up in instance attrs
+        if (
+            isinstance(node, ast.Attribute)
+            and isinstance(node.value, ast.Name)
+            and node.value.id == "self"
+        ):
+            attr_name = node.attr
+            if attr_name in self._instance_attrs:
+                return self._parse_retry_policy(self._instance_attrs[attr_name])
+            return None
+
+        if not isinstance(node, ast.Call):
+            return None
+
+        # Check if it's a RetryPolicy call
+        func_name = None
+        if isinstance(node.func, ast.Name):
+            func_name = node.func.id
+        elif isinstance(node.func, ast.Attribute):
+            func_name = node.func.attr
+
+        if func_name != "RetryPolicy":
+            return None
+
+        policy = ir.RetryPolicy()
+
+        for kw in node.keywords:
+            if kw.arg == "attempts" and isinstance(kw.value, ast.Constant):
+                # attempts means total executions, max_retries means retries after first attempt
+                # So attempts=1 -> max_retries=0 (no retries), attempts=3 -> max_retries=2
+                policy.max_retries = kw.value.value - 1
+            elif kw.arg == "exception_types" and isinstance(kw.value, ast.List):
+                for elt in kw.value.elts:
+                    if isinstance(elt, ast.Constant) and isinstance(elt.value, str):
+                        policy.exception_types.append(elt.value)
+            elif kw.arg == "backoff_seconds" and isinstance(kw.value, ast.Constant):
+                policy.backoff.seconds = int(kw.value.value)
+
+        return policy
+
+    def _parse_timeout_policy(self, node: ast.expr) -> Optional[ir.TimeoutPolicy]:
+        """Parse a timeout value into IR.
+
+        Supports:
+        - timeout=60 (int seconds)
+        - timeout=30.5 (float seconds)
+        - timeout=timedelta(seconds=30)
+        - timeout=timedelta(minutes=2)
+        - self.timeout (instance attribute reference)
+        """
+        # Handle self.attr pattern - look up in instance attrs
+        if (
+            isinstance(node, ast.Attribute)
+            and isinstance(node.value, ast.Name)
+            and node.value.id == "self"
+        ):
+            attr_name = node.attr
+            if attr_name in self._instance_attrs:
+                return self._parse_timeout_policy(self._instance_attrs[attr_name])
+            return None
+
+        policy = ir.TimeoutPolicy()
+
+        # Direct numeric value (seconds)
+        if isinstance(node, ast.Constant) and isinstance(node.value, (int, float)):
+            policy.timeout.seconds = int(node.value)
+            return policy
+
+        # timedelta(...) call
+        if isinstance(node, ast.Call):
+            func_name = None
+            if isinstance(node.func, ast.Name):
+                func_name = node.func.id
+            elif isinstance(node.func, ast.Attribute):
+                func_name = node.func.attr
+
+            if func_name == "timedelta":
+                total_seconds = 0
+                for kw in node.keywords:
+                    if isinstance(kw.value, ast.Constant):
+                        val = kw.value.value
+                        if kw.arg == "seconds":
+                            total_seconds += int(val)
+                        elif kw.arg == "minutes":
+                            total_seconds += int(val) * 60
+                        elif kw.arg == "hours":
+                            total_seconds += int(val) * 3600
+                        elif kw.arg == "days":
+                            total_seconds += int(val) * 86400
+                policy.timeout.seconds = total_seconds
+                return policy
+
+        return None
+
+    def _is_asyncio_sleep_call(self, node: ast.Call) -> bool:
+        """Check if this is an asyncio.sleep(...) call.
+
+        Supports both patterns:
+        - import asyncio; asyncio.sleep(1)
+        - from asyncio import sleep; sleep(1)
+        - from asyncio import sleep as s; s(1)
+        """
+        if isinstance(node.func, ast.Attribute):
+            # asyncio.sleep(...) pattern
+            if node.func.attr == "sleep" and isinstance(node.func.value, ast.Name):
+                return node.func.value.id == "asyncio"
+        elif isinstance(node.func, ast.Name):
+            # sleep(...) pattern - check if it's imported from asyncio
+            func_name = node.func.id
+            if func_name in self._imported_names:
+                imported = self._imported_names[func_name]
+                return imported.module == "asyncio" and imported.original_name == "sleep"
+        return False
+
+    def _convert_asyncio_sleep_to_action(self, node: ast.Call) -> ir.ActionCall:
+        """Convert asyncio.sleep(duration) to @sleep(duration=X) action call.
+
+        This creates a built-in sleep action that the scheduler handles as a
+        durable sleep - stored in the DB with a future scheduled_at time.
+        """
+        action_call = ir.ActionCall(action_name="sleep")
+
+        # Extract duration argument (positional or keyword)
+        if node.args:
+            # asyncio.sleep(1) - positional
+            expr = self._expr_to_ir_with_model_coercion(node.args[0])
+            if expr:
+                action_call.kwargs.append(ir.Kwarg(name="duration", value=expr))
+        elif node.keywords:
+            # asyncio.sleep(seconds=1) - keyword (less common)
+            for kw in node.keywords:
+                if kw.arg in ("seconds", "delay", "duration"):
+                    expr = self._expr_to_ir_with_model_coercion(kw.value)
+                    if expr:
+                        action_call.kwargs.append(ir.Kwarg(name="duration", value=expr))
+                    break
+
+        return action_call
+
+    def _is_asyncio_gather_call(self, node: ast.Call) -> bool:
+        """Check if this is an asyncio.gather(...) call.
+
+        Supports both patterns:
+        - import asyncio; asyncio.gather(a(), b())
+        - from asyncio import gather; gather(a(), b())
+        - from asyncio import gather as g; g(a(), b())
+        """
+        if isinstance(node.func, ast.Attribute):
+            # asyncio.gather(...) pattern
+            if node.func.attr == "gather" and isinstance(node.func.value, ast.Name):
+                return node.func.value.id == "asyncio"
+        elif isinstance(node.func, ast.Name):
+            # gather(...) pattern - check if it's imported from asyncio
+            func_name = node.func.id
+            if func_name in self._imported_names:
+                imported = self._imported_names[func_name]
+                return imported.module == "asyncio" and imported.original_name == "gather"
+        return False
+
+    def _convert_asyncio_gather(
+        self, node: ast.Call
+    ) -> Optional[Union[ir.ParallelExpr, ir.SpreadExpr]]:
+        """Convert asyncio.gather(...) to ParallelExpr or SpreadExpr IR.
+
+        Handles two patterns:
+        1. Static gather: asyncio.gather(a(), b(), c()) -> ParallelExpr
+        2. Spread gather: asyncio.gather(*[action(x) for x in items]) -> SpreadExpr
+
+        Args:
+            node: The asyncio.gather() Call node
+
+        Returns:
+            A ParallelExpr, SpreadExpr, or None if conversion fails.
+        """
+        return_exceptions_value: Optional[ast.expr] = None
+        for kw in node.keywords:
+            if kw.arg is None:
+                line = node.lineno if hasattr(node, "lineno") else None
+                col = node.col_offset if hasattr(node, "col_offset") else None
+                raise UnsupportedPatternError(
+                    "asyncio.gather() must be called with return_exceptions=True",
+                    RECOMMENDATIONS["gather_return_exceptions"],
+                    line=line,
+                    col=col,
+                )
+            if kw.arg == "return_exceptions":
+                return_exceptions_value = kw.value
+                break
+
+        if not (
+            isinstance(return_exceptions_value, ast.Constant)
+            and return_exceptions_value.value is True
+        ):
+            line = node.lineno if hasattr(node, "lineno") else None
+            col = node.col_offset if hasattr(node, "col_offset") else None
+            raise UnsupportedPatternError(
+                "asyncio.gather() must be called with return_exceptions=True",
+                RECOMMENDATIONS["gather_return_exceptions"],
+                line=line,
+                col=col,
+            )
+
+        # Check for starred expressions - spread pattern
+        if len(node.args) == 1 and isinstance(node.args[0], ast.Starred):
+            starred = node.args[0]
+            # Only list comprehensions are supported for spread
+            if isinstance(starred.value, ast.ListComp):
+                return self._convert_listcomp_to_spread_expr(starred.value)
+            else:
+                # Spreading a variable or other expression is not supported
+                line = getattr(node, "lineno", None)
+                col = getattr(node, "col_offset", None)
+                if isinstance(starred.value, ast.Name):
+                    var_name = starred.value.id
+                    raise UnsupportedPatternError(
+                        f"Spreading variable '{var_name}' in asyncio.gather() is not supported",
+                        RECOMMENDATIONS["gather_variable_spread"],
+                        line=line,
+                        col=col,
+                    )
+                else:
+                    raise UnsupportedPatternError(
+                        "Spreading non-list-comprehension expressions in asyncio.gather() is not supported",
+                        RECOMMENDATIONS["gather_variable_spread"],
+                        line=line,
+                        col=col,
+                    )
+
+        # Standard case: gather(a(), b(), c()) -> ParallelExpr
+        parallel = ir.ParallelExpr()
+
+        # Each argument to gather() should be an action call
+        for arg in node.args:
+            call = self._convert_gather_arg_to_call(arg)
+            if call:
+                parallel.calls.append(call)
+
+        # Only return if we have calls
+        if not parallel.calls:
+            return None
+
+        return parallel
+
+    def _convert_listcomp_to_spread_expr(self, listcomp: ast.ListComp) -> Optional[ir.SpreadExpr]:
+        """Convert a list comprehension to SpreadExpr IR.
+
+        Handles patterns like:
+            [action(x=item) for item in collection]
+            [self.run_action(action(x=item), retry=..., timeout=...) for item in collection]
+
+        The comprehension must have exactly one generator with no conditions,
+        and the element must be an action call (optionally wrapped in run_action).
+
+        Args:
+            listcomp: The ListComp AST node
+
+        Returns:
+            A SpreadExpr, or None if conversion fails.
+        """
+        # Only support simple list comprehensions with one generator
+        if len(listcomp.generators) != 1:
+            line = getattr(listcomp, "lineno", None)
+            col = getattr(listcomp, "col_offset", None)
+            raise UnsupportedPatternError(
+                "Spread pattern only supports a single loop variable",
+                "Use a simple list comprehension: [action(x) for x in items]",
+                line=line,
+                col=col,
+            )
+
+        gen = listcomp.generators[0]
+
+        # Check for conditions - not supported
+        if gen.ifs:
+            line = getattr(listcomp, "lineno", None)
+            col = getattr(listcomp, "col_offset", None)
+            raise UnsupportedPatternError(
+                "Spread pattern does not support conditions in list comprehension",
+                "Remove the 'if' clause from the comprehension",
+                line=line,
+                col=col,
+            )
+
+        # Get the loop variable name
+        if not isinstance(gen.target, ast.Name):
+            line = getattr(listcomp, "lineno", None)
+            col = getattr(listcomp, "col_offset", None)
+            raise UnsupportedPatternError(
+                "Spread pattern requires a simple loop variable",
+                "Use a simple variable: [action(x) for x in items]",
+                line=line,
+                col=col,
+            )
+        loop_var = gen.target.id
+
+        # Get the collection expression
+        collection_expr = self._expr_to_ir_with_model_coercion(gen.iter)
+        if not collection_expr:
+            line = getattr(listcomp, "lineno", None)
+            col = getattr(listcomp, "col_offset", None)
+            raise UnsupportedPatternError(
+                "Could not convert collection expression in spread pattern",
+                "Ensure the collection is a simple variable or expression",
+                line=line,
+                col=col,
+            )
+
+        # The element must be a call (either action call or run_action wrapper)
+        if not isinstance(listcomp.elt, ast.Call):
+            line = getattr(listcomp, "lineno", None)
+            col = getattr(listcomp, "col_offset", None)
+            raise UnsupportedPatternError(
+                "Spread pattern requires an action call in the list comprehension",
+                "Use: [action(x=item) for item in items]",
+                line=line,
+                col=col,
+            )
+
+        # Check for self.run_action(...) wrapper pattern
+        action_call: Optional[ir.ActionCall] = None
+        if self._is_run_action_call(listcomp.elt):
+            # Extract the inner action call from run_action's first argument
+            if listcomp.elt.args:
+                inner_call = listcomp.elt.args[0]
+                if isinstance(inner_call, ast.Call):
+                    action_call = self._extract_action_call_from_call(inner_call)
+                    if action_call:
+                        # Extract policies (retry, timeout) from run_action kwargs
+                        self._extract_policies_from_run_action(listcomp.elt, action_call)
+        else:
+            # Direct action call
+            action_call = self._extract_action_call_from_call(listcomp.elt)
+
+        if not action_call:
+            line = getattr(listcomp, "lineno", None)
+            col = getattr(listcomp, "col_offset", None)
+            raise UnsupportedPatternError(
+                "Spread pattern element must be an @action call",
+                "Ensure the function is decorated with @action, or use self.run_action(action(...), ...)",
+                line=line,
+                col=col,
+            )
+
+        # Build the SpreadExpr
+        spread = ir.SpreadExpr()
+        spread.collection.CopyFrom(collection_expr)
+        spread.loop_var = loop_var
+        spread.action.CopyFrom(action_call)
+
+        return spread
+
+    def _convert_gather_arg_to_call(self, node: ast.expr) -> Optional[ir.Call]:
+        """Convert a gather argument to an IR Call.
+
+        Handles both action calls and regular function calls.
+        """
+        if not isinstance(node, ast.Call):
+            return None
+
+        # Try to extract as an action call first
+        action_call = self._extract_action_call_from_call(node)
+        if action_call:
+            call = ir.Call()
+            call.action.CopyFrom(action_call)
+            return call
+
+        # Fall back to regular function call
+        func_call = self._convert_to_function_call(node)
+        if func_call:
+            call = ir.Call()
+            call.function.CopyFrom(func_call)
+            return call
+
+        return None
+
+    def _convert_to_function_call(self, node: ast.Call) -> Optional[ir.FunctionCall]:
+        """Convert an AST Call to IR FunctionCall."""
+        func_name = self._get_func_name(node.func)
+        if not func_name:
+            return None
+
+        fn_call = ir.FunctionCall(name=func_name)
+        global_function = _global_function_for_call(func_name, node)
+        if global_function is not None:
+            fn_call.global_function = global_function
+
+        # Add positional args
+        for arg in node.args:
+            expr = self._expr_to_ir_with_model_coercion(arg)
+            if expr:
+                fn_call.args.append(expr)
+
+        # Add keyword args
+        for kw in node.keywords:
+            if kw.arg:
+                expr = self._expr_to_ir_with_model_coercion(kw.value)
+                if expr:
+                    fn_call.kwargs.append(ir.Kwarg(name=kw.arg, value=expr))
+
+        # Fill in missing kwargs with default values from function signature
+        self._fill_default_kwargs_for_expr(fn_call)
+
+        return fn_call
+
+    def _get_func_name(self, node: ast.expr) -> Optional[str]:
+        """Get function name from a func node."""
+        if isinstance(node, ast.Name):
+            return node.id
+        elif isinstance(node, ast.Attribute):
+            # Handle chained attributes like obj.method
+            parts = []
+            current = node
+            while isinstance(current, ast.Attribute):
+                parts.append(current.attr)
+                current = current.value
+            if isinstance(current, ast.Name):
+                parts.append(current.id)
+            name = ".".join(reversed(parts))
+            if name.startswith("self."):
+                return name[5:]
+            return name
+        return None
+
+    def _convert_model_constructor_if_needed(self, node: ast.Call) -> Optional[ir.Expr]:
+        model_name = self._is_model_constructor(node)
+        if model_name:
+            return self._convert_model_constructor_to_dict(node, model_name)
+        return None
+
+    def _resolve_enum_attribute(self, node: ast.Attribute) -> Optional[ir.Expr]:
+        value = _resolve_enum_attribute_value(node, self._module_globals)
+        if value is None:
+            return None
+        literal = _constant_to_literal(value)
+        if literal is None:
+            line = getattr(node, "lineno", None)
+            col = getattr(node, "col_offset", None)
+            raise UnsupportedPatternError(
+                "Enum value must be a primitive literal",
+                RECOMMENDATIONS["unsupported_literal"],
+                line=line,
+                col=col,
+            )
+        expr = ir.Expr(span=_make_span(node))
+        expr.literal.CopyFrom(literal)
+        return expr
+
+    def _is_exception_class(self, class_name: str) -> bool:
+        """Check if a class name refers to an exception class.
+
+        This checks both built-in exceptions and imported exception classes.
+        """
+        # Check built-in exceptions first
+        builtin_exc = (
+            getattr(__builtins__, class_name, None)
+            if isinstance(__builtins__, dict)
+            else getattr(__builtins__, class_name, None)
+        )
+        if builtin_exc is None:
+            # Try getting from builtins module directly
+            import builtins
+
+            builtin_exc = getattr(builtins, class_name, None)
+        if (
+            builtin_exc is not None
+            and isinstance(builtin_exc, type)
+            and issubclass(builtin_exc, BaseException)
+        ):
+            return True
+
+        # Check module globals for imported exception classes
+        cls = self._module_globals.get(class_name)
+        if cls is not None and isinstance(cls, type) and issubclass(cls, BaseException):
+            return True
+
+        return False
+
+    def _expr_to_ir_with_model_coercion(self, node: ast.expr) -> Optional[ir.Expr]:
+        """Convert an AST expression to IR, converting model constructors to dicts."""
+        result = _expr_to_ir(
+            node,
+            model_converter=self._convert_model_constructor_if_needed,
+            enum_resolver=self._resolve_enum_attribute,
+            exception_class_resolver=self._is_exception_class,
+        )
+        # Post-process to fill in default kwargs for function calls (recursively)
+        if result is not None:
+            self._fill_default_kwargs_recursive(result)
+        return result
+
+    def _fill_default_kwargs_recursive(self, expr: ir.Expr) -> None:
+        """Recursively fill in default kwargs for all function calls in an expression."""
+        if expr.HasField("function_call"):
+            self._fill_default_kwargs_for_expr(expr.function_call)
+            # Recurse into function call args and kwargs
+            for arg in expr.function_call.args:
+                self._fill_default_kwargs_recursive(arg)
+            for kwarg in expr.function_call.kwargs:
+                if kwarg.value:
+                    self._fill_default_kwargs_recursive(kwarg.value)
+        elif expr.HasField("binary_op"):
+            if expr.binary_op.left:
+                self._fill_default_kwargs_recursive(expr.binary_op.left)
+            if expr.binary_op.right:
+                self._fill_default_kwargs_recursive(expr.binary_op.right)
+        elif expr.HasField("unary_op"):
+            if expr.unary_op.operand:
+                self._fill_default_kwargs_recursive(expr.unary_op.operand)
+        elif expr.HasField("list"):
+            for elem in expr.list.elements:
+                self._fill_default_kwargs_recursive(elem)
+        elif expr.HasField("dict"):
+            for entry in expr.dict.entries:
+                if entry.key:
+                    self._fill_default_kwargs_recursive(entry.key)
+                if entry.value:
+                    self._fill_default_kwargs_recursive(entry.value)
+        elif expr.HasField("index"):
+            if expr.index.object:
+                self._fill_default_kwargs_recursive(expr.index.object)
+            if expr.index.index:
+                self._fill_default_kwargs_recursive(expr.index.index)
+        elif expr.HasField("dot"):
+            if expr.dot.object:
+                self._fill_default_kwargs_recursive(expr.dot.object)
+
+    def _fill_default_kwargs_for_expr(self, fn_call: ir.FunctionCall) -> None:
+        """Fill in missing kwargs with default values from the function signature."""
+        sig_info = self._ctx.function_signatures.get(fn_call.name)
+        if sig_info is None:
+            return
+
+        # Track which parameters are already provided
+        provided_by_position: Set[str] = set()
+        provided_by_kwarg: Set[str] = set()
+
+        # Positional args map to parameters in order
+        for idx, _arg in enumerate(fn_call.args):
+            if idx < len(sig_info.parameters):
+                provided_by_position.add(sig_info.parameters[idx].name)
+
+        # Kwargs are named
+        for kwarg in fn_call.kwargs:
+            provided_by_kwarg.add(kwarg.name)
+
+        # Add defaults for missing parameters
+        for param in sig_info.parameters:
+            if param.name in provided_by_position or param.name in provided_by_kwarg:
+                continue
+            if not param.has_default:
+                continue
+
+            # Convert the default value to an IR expression
+            literal = _constant_to_literal(param.default_value)
+            if literal is not None:
+                expr = ir.Expr()
+                expr.literal.CopyFrom(literal)
+                fn_call.kwargs.append(ir.Kwarg(name=param.name, value=expr))
+
+    def _extract_action_call_from_awaitable(self, node: ast.expr) -> Optional[ir.ActionCall]:
+        """Extract action call from an awaitable expression."""
+        if isinstance(node, ast.Call):
+            return self._extract_action_call_from_call(node)
+        return None
+
+    def _extract_action_call_from_call(self, node: ast.Call) -> Optional[ir.ActionCall]:
+        """Extract action call info from a Call node.
+
+        Converts positional arguments to keyword arguments using the action's
+        signature introspection. This ensures all arguments are named in the IR.
+
+        Pydantic models and dataclass constructors passed as arguments are
+        automatically converted to dict expressions.
+        """
+        action_name = self._get_action_name(node.func)
+        if not action_name:
+            return None
+
+        if action_name not in self._action_defs:
+            return None
+
+        action_def = self._action_defs[action_name]
+        action_call = ir.ActionCall(action_name=action_def.action_name)
+
+        # Set the module name so the worker knows where to find the action
+        if action_def.module_name:
+            action_call.module_name = action_def.module_name
+
+        # Get parameter names from signature for positional arg conversion
+        param_names = list(action_def.signature.parameters.keys())
+
+        # Convert positional args to kwargs using signature introspection
+        # Model constructors are converted to dict expressions
+        for i, arg in enumerate(node.args):
+            if i < len(param_names):
+                expr = self._expr_to_ir_with_model_coercion(arg)
+                if expr:
+                    kwarg = ir.Kwarg(name=param_names[i], value=expr)
+                    action_call.kwargs.append(kwarg)
+
+        # Add explicit kwargs
+        # Model constructors are converted to dict expressions
+        for kw in node.keywords:
+            if kw.arg:
+                expr = self._expr_to_ir_with_model_coercion(kw.value)
+                if expr:
+                    kwarg = ir.Kwarg(name=kw.arg, value=expr)
+                    action_call.kwargs.append(kwarg)
+
+        return action_call
+
+    def _get_action_name(self, func: ast.expr) -> Optional[str]:
+        """Get the action name from a function expression."""
+        if isinstance(func, ast.Name):
+            return func.id
+        elif isinstance(func, ast.Attribute):
+            return func.attr
+        return None
+
+    def _get_assign_target(self, targets: List[ast.expr]) -> Optional[str]:
+        """Get the target variable name from assignment targets (single target only)."""
+        if targets and isinstance(targets[0], ast.Name):
+            return targets[0].id
+        return None
+
+    def _get_assign_targets(self, targets: List[ast.expr]) -> List[str]:
+        """Get all target variable names from assignment targets (including tuple unpacking)."""
+        result: List[str] = []
+        for t in targets:
+            if isinstance(t, ast.Name):
+                result.append(t.id)
+            elif isinstance(t, ast.Subscript):
+                formatted = _format_subscript_target(t)
+                if formatted:
+                    result.append(formatted)
+            elif isinstance(t, ast.Tuple):
+                for elt in t.elts:
+                    if isinstance(elt, ast.Name):
+                        result.append(elt.id)
+        return result
+
+
+def _make_span(node: ast.AST) -> ir.Span:
+    """Create a Span from an AST node."""
+    return ir.Span(
+        start_line=getattr(node, "lineno", 0),
+        start_col=getattr(node, "col_offset", 0),
+        end_line=getattr(node, "end_lineno", 0) or 0,
+        end_col=getattr(node, "end_col_offset", 0) or 0,
+    )
+
+
+def _attribute_chain(node: ast.Attribute) -> Optional[List[str]]:
+    parts: List[str] = []
+    current: ast.AST = node
+    while isinstance(current, ast.Attribute):
+        parts.append(current.attr)
+        current = current.value
+    if isinstance(current, ast.Name):
+        parts.append(current.id)
+        return list(reversed(parts))
+    return None
+
+
+def _resolve_enum_attribute_value(
+    node: ast.Attribute,
+    module_globals: Mapping[str, Any],
+) -> Optional[Any]:
+    chain = _attribute_chain(node)
+    if not chain or len(chain) < 2:
+        return None
+
+    current = module_globals.get(chain[0])
+    if current is None:
+        return None
+
+    for part in chain[1:-1]:
+        try:
+            current_dict = current.__dict__
+        except AttributeError:
+            return None
+        current = current_dict.get(part)
+        if current is None:
+            return None
+
+    member_name = chain[-1]
+    if isinstance(current, EnumMeta):
+        member = current.__members__.get(member_name)
+        if member is None:
+            return None
+        return member.value
+
+    return None
+
+
+def _try_convert_isinstance_to_isexception(
+    expr: ast.Call,
+    exception_class_resolver: Callable[[str], bool],
+    model_converter: Optional[Callable[[ast.Call], Optional[ir.Expr]]] = None,
+    enum_resolver: Optional[Callable[[ast.Attribute], Optional[ir.Expr]]] = None,
+) -> Optional[ir.Expr]:
+    """Try to convert isinstance(x, ExceptionClass) to isexception(x, "ExceptionClass").
+
+    Returns None if this is not an isinstance call or if the class is not an exception.
+    Raises UnsupportedPatternError if isinstance is used with a non-exception class.
+    """
+    # Check if this is an isinstance call
+    func_name = _get_func_name(expr.func)
+    if func_name != "isinstance":
+        return None
+
+    # isinstance requires exactly 2 positional arguments
+    if len(expr.args) != 2 or expr.keywords:
+        line = expr.lineno if hasattr(expr, "lineno") else None
+        col = expr.col_offset if hasattr(expr, "col_offset") else None
+        raise UnsupportedPatternError(
+            "isinstance() requires exactly 2 positional arguments",
+            RECOMMENDATIONS["sync_function_call"],
+            line=line,
+            col=col,
+        )
+
+    # Extract exception class names from the second argument
+    second_arg = expr.args[1]
+    exception_names: List[str] = []
+
+    if isinstance(second_arg, ast.Name):
+        # Single class: isinstance(x, ValueError)
+        class_name = second_arg.id
+        if not exception_class_resolver(class_name):
+            line = expr.lineno if hasattr(expr, "lineno") else None
+            col = expr.col_offset if hasattr(expr, "col_offset") else None
+            raise UnsupportedPatternError(
+                f"isinstance() with non-exception class '{class_name}' is not supported in workflows",
+                "isinstance() can only be used to check exception types in workflow code. "
+                "Move other type checks to an @action.",
+                line=line,
+                col=col,
+            )
+        exception_names.append(class_name)
+    elif isinstance(second_arg, ast.Tuple):
+        # Tuple of classes: isinstance(x, (ValueError, TypeError))
+        for elt in second_arg.elts:
+            if not isinstance(elt, ast.Name):
+                line = expr.lineno if hasattr(expr, "lineno") else None
+                col = expr.col_offset if hasattr(expr, "col_offset") else None
+                raise UnsupportedPatternError(
+                    "isinstance() class argument must be a simple name or tuple of names",
+                    "Use simple class names like ValueError or (ValueError, TypeError).",
+                    line=line,
+                    col=col,
+                )
+            class_name = elt.id
+            if not exception_class_resolver(class_name):
+                line = expr.lineno if hasattr(expr, "lineno") else None
+                col = expr.col_offset if hasattr(expr, "col_offset") else None
+                raise UnsupportedPatternError(
+                    f"isinstance() with non-exception class '{class_name}' is not supported in workflows",
+                    "isinstance() can only be used to check exception types in workflow code. "
+                    "Move other type checks to an @action.",
+                    line=line,
+                    col=col,
+                )
+            exception_names.append(class_name)
+    else:
+        line = expr.lineno if hasattr(expr, "lineno") else None
+        col = expr.col_offset if hasattr(expr, "col_offset") else None
+        raise UnsupportedPatternError(
+            "isinstance() class argument must be a simple name or tuple of names",
+            "Use simple class names like ValueError or (ValueError, TypeError).",
+            line=line,
+            col=col,
+        )
+
+    # Convert the first argument (the value being checked) to IR
+    value_expr = _expr_to_ir(
+        expr.args[0],
+        model_converter=model_converter,
+        enum_resolver=enum_resolver,
+        exception_class_resolver=exception_class_resolver,
+    )
+    if not value_expr:
+        return None
+
+    # Build the isexception call
+    # If single exception: isexception(x, "ValueError")
+    # If multiple: isexception(x, ["ValueError", "TypeError"])
+    result = ir.Expr(span=_make_span(expr))
+
+    if len(exception_names) == 1:
+        # Single exception name as string
+        type_arg = ir.Expr(span=_make_span(second_arg))
+        type_arg.literal.CopyFrom(ir.Literal(string_value=exception_names[0]))
+        args = [value_expr, type_arg]
+    else:
+        # Multiple exception names as list of strings
+        type_elements = []
+        for name in exception_names:
+            elem = ir.Expr()
+            elem.literal.CopyFrom(ir.Literal(string_value=name))
+            type_elements.append(elem)
+        type_arg = ir.Expr(span=_make_span(second_arg))
+        type_arg.list.CopyFrom(ir.ListExpr(elements=type_elements))
+        args = [value_expr, type_arg]
+
+    func_call = ir.FunctionCall(
+        name="isexception",
+        args=args,
+        kwargs=[],
+    )
+    func_call.global_function = ir.GlobalFunction.GLOBAL_FUNCTION_ISEXCEPTION
+    result.function_call.CopyFrom(func_call)
+    return result
+
+
+def _expr_to_ir(
+    expr: ast.AST,
+    model_converter: Optional[Callable[[ast.Call], Optional[ir.Expr]]] = None,
+    enum_resolver: Optional[Callable[[ast.Attribute], Optional[ir.Expr]]] = None,
+    exception_class_resolver: Optional[Callable[[str], bool]] = None,
+) -> Optional[ir.Expr]:
+    """Convert Python AST expression to IR Expr.
+
+    Args:
+        expr: The AST expression to convert.
+        model_converter: Optional callback to convert model constructors.
+        enum_resolver: Optional callback to resolve enum attributes.
+        exception_class_resolver: Optional callback that takes a class name and
+            returns True if it's an exception class. Used to transform
+            isinstance(x, ExceptionClass) to isexception(x, "ExceptionClass").
+    """
+    result = ir.Expr(span=_make_span(expr))
+
+    if isinstance(expr, ast.Call) and model_converter:
+        converted = model_converter(expr)
+        if converted:
+            return converted
+
+    # Handle isinstance(x, ExceptionClass) -> isexception(x, "ExceptionClass")
+    if isinstance(expr, ast.Call) and exception_class_resolver:
+        isinstance_result = _try_convert_isinstance_to_isexception(
+            expr, exception_class_resolver, model_converter, enum_resolver
+        )
+        if isinstance_result is not None:
+            return isinstance_result
+
+    if isinstance(expr, ast.Name):
+        result.variable.CopyFrom(ir.Variable(name=expr.id))
+        return result
+
+    if isinstance(expr, ast.Constant):
+        literal = _constant_to_literal(expr.value)
+        if literal:
+            result.literal.CopyFrom(literal)
+            return result
+
+    if isinstance(expr, ast.BinOp):
+        left = _expr_to_ir(
+            expr.left,
+            model_converter=model_converter,
+            enum_resolver=enum_resolver,
+            exception_class_resolver=exception_class_resolver,
+        )
+        right = _expr_to_ir(
+            expr.right,
+            model_converter=model_converter,
+            enum_resolver=enum_resolver,
+            exception_class_resolver=exception_class_resolver,
+        )
+        op = _bin_op_to_ir(expr.op)
+        if left and right and op:
+            result.binary_op.CopyFrom(ir.BinaryOp(left=left, op=op, right=right))
+            return result
+
+    if isinstance(expr, ast.UnaryOp):
+        operand = _expr_to_ir(
+            expr.operand,
+            model_converter=model_converter,
+            enum_resolver=enum_resolver,
+            exception_class_resolver=exception_class_resolver,
+        )
+        op = _unary_op_to_ir(expr.op)
+        if operand and op:
+            result.unary_op.CopyFrom(ir.UnaryOp(op=op, operand=operand))
+            return result
+
+    if isinstance(expr, ast.Compare):
+        left = _expr_to_ir(
+            expr.left,
+            model_converter=model_converter,
+            enum_resolver=enum_resolver,
+            exception_class_resolver=exception_class_resolver,
+        )
+        if not left:
+            return None
+        # For simplicity, handle single comparison
+        if expr.ops and expr.comparators:
+            op = _cmp_op_to_ir(expr.ops[0])
+            right = _expr_to_ir(
+                expr.comparators[0],
+                model_converter=model_converter,
+                enum_resolver=enum_resolver,
+            )
+            if op and right:
+                result.binary_op.CopyFrom(ir.BinaryOp(left=left, op=op, right=right))
+                return result
+
+    if isinstance(expr, ast.BoolOp):
+        values = [
+            _expr_to_ir(
+                v,
+                model_converter=model_converter,
+                enum_resolver=enum_resolver,
+                exception_class_resolver=exception_class_resolver,
+            )
+            for v in expr.values
+        ]
+        if all(v for v in values):
+            op = _bool_op_to_ir(expr.op)
+            if op and len(values) >= 2:
+                # Chain boolean ops: a and b and c -> (a and b) and c
+                result_expr = values[0]
+                for v in values[1:]:
+                    if result_expr and v:
+                        new_result = ir.Expr()
+                        new_result.binary_op.CopyFrom(ir.BinaryOp(left=result_expr, op=op, right=v))
+                        result_expr = new_result
+                return result_expr
+
+    if isinstance(expr, ast.List):
+        elements = [
+            _expr_to_ir(
+                e,
+                model_converter=model_converter,
+                enum_resolver=enum_resolver,
+                exception_class_resolver=exception_class_resolver,
+            )
+            for e in expr.elts
+        ]
+        if all(e for e in elements):
+            list_expr = ir.ListExpr(elements=[e for e in elements if e])
+            result.list.CopyFrom(list_expr)
+            return result
+
+    if isinstance(expr, ast.Dict):
+        entries: List[ir.DictEntry] = []
+        for k, v in zip(expr.keys, expr.values, strict=False):
+            if k:
+                key_expr = _expr_to_ir(
+                    k,
+                    model_converter=model_converter,
+                    enum_resolver=enum_resolver,
+                )
+                value_expr = _expr_to_ir(
+                    v,
+                    model_converter=model_converter,
+                    enum_resolver=enum_resolver,
+                )
+                if key_expr and value_expr:
+                    entries.append(ir.DictEntry(key=key_expr, value=value_expr))
+        result.dict.CopyFrom(ir.DictExpr(entries=entries))
+        return result
+
+    if isinstance(expr, ast.Subscript):
+        obj = _expr_to_ir(
+            expr.value,
+            model_converter=model_converter,
+            enum_resolver=enum_resolver,
+            exception_class_resolver=exception_class_resolver,
+        )
+        index = (
+            _expr_to_ir(
+                expr.slice,
+                model_converter=model_converter,
+                enum_resolver=enum_resolver,
+            )
+            if isinstance(expr.slice, ast.AST)
+            else None
+        )
+        if obj and index:
+            result.index.CopyFrom(ir.IndexAccess(object=obj, index=index))
+            return result
+
+    if isinstance(expr, ast.Attribute):
+        if enum_resolver:
+            resolved = enum_resolver(expr)
+            if resolved:
+                return resolved
+        obj = _expr_to_ir(
+            expr.value,
+            model_converter=model_converter,
+            enum_resolver=enum_resolver,
+            exception_class_resolver=exception_class_resolver,
+        )
+        if obj:
+            result.dot.CopyFrom(ir.DotAccess(object=obj, attribute=expr.attr))
+            return result
+
+    if isinstance(expr, ast.Await) and isinstance(expr.value, ast.Call):
+        func_name = _get_func_name(expr.value.func)
+        if func_name:
+            args = [
+                _expr_to_ir(
+                    a,
+                    model_converter=model_converter,
+                    enum_resolver=enum_resolver,
+                )
+                for a in expr.value.args
+            ]
+            kwargs: List[ir.Kwarg] = []
+            for kw in expr.value.keywords:
+                if kw.arg:
+                    kw_expr = _expr_to_ir(
+                        kw.value,
+                        model_converter=model_converter,
+                        enum_resolver=enum_resolver,
+                    )
+                    if kw_expr:
+                        kwargs.append(ir.Kwarg(name=kw.arg, value=kw_expr))
+            func_call = ir.FunctionCall(
+                name=func_name,
+                args=[a for a in args if a],
+                kwargs=kwargs,
+            )
+            global_function = _global_function_for_call(func_name, expr.value)
+            if global_function is not None:
+                func_call.global_function = global_function
+            result.function_call.CopyFrom(func_call)
+            return result
+
+    if isinstance(expr, ast.Call):
+        # Function call
+        if not _is_self_method_call(expr):
+            func_name = _get_func_name(expr.func) or "unknown"
+            if isinstance(expr.func, ast.Attribute):
+                line = expr.lineno if hasattr(expr, "lineno") else None
+                col = expr.col_offset if hasattr(expr, "col_offset") else None
+                raise UnsupportedPatternError(
+                    f"Calling synchronous function '{func_name}()' directly is not supported",
+                    RECOMMENDATIONS["sync_function_call"],
+                    line=line,
+                    col=col,
+                )
+            if func_name not in ALLOWED_SYNC_FUNCTIONS:
+                line = expr.lineno if hasattr(expr, "lineno") else None
+                col = expr.col_offset if hasattr(expr, "col_offset") else None
+                raise UnsupportedPatternError(
+                    f"Calling synchronous function '{func_name}()' directly is not supported",
+                    RECOMMENDATIONS["sync_function_call"],
+                    line=line,
+                    col=col,
+                )
+        func_name = _get_func_name(expr.func)
+        if func_name:
+            args = [
+                _expr_to_ir(
+                    a,
+                    model_converter=model_converter,
+                    enum_resolver=enum_resolver,
+                )
+                for a in expr.args
+            ]
+            kwargs: List[ir.Kwarg] = []
+            for kw in expr.keywords:
+                if kw.arg:
+                    kw_expr = _expr_to_ir(
+                        kw.value,
+                        model_converter=model_converter,
+                        enum_resolver=enum_resolver,
+                    )
+                    if kw_expr:
+                        kwargs.append(ir.Kwarg(name=kw.arg, value=kw_expr))
+            func_call = ir.FunctionCall(
+                name=func_name,
+                args=[a for a in args if a],
+                kwargs=kwargs,
+            )
+            global_function = _global_function_for_call(func_name, expr)
+            if global_function is not None:
+                func_call.global_function = global_function
+            result.function_call.CopyFrom(func_call)
+            return result
+
+    if isinstance(expr, ast.Tuple):
+        # Handle tuple as list for now
+        elements = [
+            _expr_to_ir(
+                e,
+                model_converter=model_converter,
+                enum_resolver=enum_resolver,
+                exception_class_resolver=exception_class_resolver,
+            )
+            for e in expr.elts
+        ]
+        if all(e for e in elements):
+            list_expr = ir.ListExpr(elements=[e for e in elements if e])
+            result.list.CopyFrom(list_expr)
+            return result
+
+    # Check for unsupported expression types
+    _check_unsupported_expression(expr)
+
+    return None
+
+
+def _check_unsupported_expression(expr: ast.AST) -> None:
+    """Check for unsupported expression types and raise descriptive errors."""
+    line = getattr(expr, "lineno", None)
+    col = getattr(expr, "col_offset", None)
+
+    if isinstance(expr, ast.Constant):
+        if _constant_to_literal(expr.value) is None:
+            raise UnsupportedPatternError(
+                f"Unsupported literal type '{type(expr.value).__name__}'",
+                RECOMMENDATIONS["unsupported_literal"],
+                line=line,
+                col=col,
+            )
+
+    if isinstance(expr, ast.JoinedStr):
+        raise UnsupportedPatternError(
+            "F-strings are not supported",
+            RECOMMENDATIONS["fstring"],
+            line=line,
+            col=col,
+        )
+    elif isinstance(expr, ast.Lambda):
+        raise UnsupportedPatternError(
+            "Lambda expressions are not supported",
+            RECOMMENDATIONS["lambda"],
+            line=line,
+            col=col,
+        )
+    elif isinstance(expr, ast.ListComp):
+        raise UnsupportedPatternError(
+            "List comprehensions are not supported in this context",
+            RECOMMENDATIONS["list_comprehension"],
+            line=line,
+            col=col,
+        )
+    elif isinstance(expr, ast.DictComp):
+        raise UnsupportedPatternError(
+            "Dict comprehensions are not supported in this context",
+            RECOMMENDATIONS["dict_comprehension"],
+            line=line,
+            col=col,
+        )
+    elif isinstance(expr, ast.SetComp):
+        raise UnsupportedPatternError(
+            "Set comprehensions are not supported",
+            RECOMMENDATIONS["set_comprehension"],
+            line=line,
+            col=col,
+        )
+    elif isinstance(expr, ast.GeneratorExp):
+        raise UnsupportedPatternError(
+            "Generator expressions are not supported",
+            RECOMMENDATIONS["generator"],
+            line=line,
+            col=col,
+        )
+    elif isinstance(expr, ast.NamedExpr):
+        raise UnsupportedPatternError(
+            "The walrus operator (:=) is not supported",
+            RECOMMENDATIONS["walrus"],
+            line=line,
+            col=col,
+        )
+    elif isinstance(expr, ast.Yield) or isinstance(expr, ast.YieldFrom):
+        raise UnsupportedPatternError(
+            "Yield expressions are not supported",
+            RECOMMENDATIONS["yield_statement"],
+            line=line,
+            col=col,
+        )
+    elif isinstance(expr, ast.expr):
+        raise UnsupportedPatternError(
+            f"Unsupported expression type '{type(expr).__name__}'",
+            RECOMMENDATIONS["unsupported_expression"],
+            line=line,
+            col=col,
+        )
+
+
+def _format_subscript_target(target: ast.Subscript) -> Optional[str]:
+    """Convert a subscript target to a string representation for tracking."""
+    if not isinstance(target.value, ast.Name):
+        return None
+
+    base = target.value.id
+    try:
+        index_str = ast.unparse(target.slice)
+    except Exception:
+        return None
+
+    return f"{base}[{index_str}]"
+
+
+def _constant_to_literal(value: Any) -> Optional[ir.Literal]:
+    """Convert a Python constant to IR Literal."""
+    literal = ir.Literal()
+    if value is None:
+        literal.is_none = True
+    elif isinstance(value, bool):
+        literal.bool_value = value
+    elif isinstance(value, int):
+        literal.int_value = value
+    elif isinstance(value, float):
+        literal.float_value = value
+    elif isinstance(value, str):
+        literal.string_value = value
+    else:
+        return None
+    return literal
+
+
+def _bin_op_to_ir(op: ast.operator) -> Optional[ir.BinaryOperator]:
+    """Convert Python binary operator to IR BinaryOperator."""
+    mapping = {
+        ast.Add: ir.BinaryOperator.BINARY_OP_ADD,
+        ast.Sub: ir.BinaryOperator.BINARY_OP_SUB,
+        ast.Mult: ir.BinaryOperator.BINARY_OP_MUL,
+        ast.Div: ir.BinaryOperator.BINARY_OP_DIV,
+        ast.FloorDiv: ir.BinaryOperator.BINARY_OP_FLOOR_DIV,
+        ast.Mod: ir.BinaryOperator.BINARY_OP_MOD,
+    }
+    return mapping.get(type(op))
+
+
+def _unary_op_to_ir(op: ast.unaryop) -> Optional[ir.UnaryOperator]:
+    """Convert Python unary operator to IR UnaryOperator."""
+    mapping = {
+        ast.USub: ir.UnaryOperator.UNARY_OP_NEG,
+        ast.Not: ir.UnaryOperator.UNARY_OP_NOT,
+    }
+    return mapping.get(type(op))
+
+
+def _cmp_op_to_ir(op: ast.cmpop) -> Optional[ir.BinaryOperator]:
+    """Convert Python comparison operator to IR BinaryOperator."""
+    mapping = {
+        ast.Eq: ir.BinaryOperator.BINARY_OP_EQ,
+        ast.NotEq: ir.BinaryOperator.BINARY_OP_NE,
+        ast.Lt: ir.BinaryOperator.BINARY_OP_LT,
+        ast.LtE: ir.BinaryOperator.BINARY_OP_LE,
+        ast.Gt: ir.BinaryOperator.BINARY_OP_GT,
+        ast.GtE: ir.BinaryOperator.BINARY_OP_GE,
+        ast.In: ir.BinaryOperator.BINARY_OP_IN,
+        ast.NotIn: ir.BinaryOperator.BINARY_OP_NOT_IN,
+    }
+    return mapping.get(type(op))
+
+
+def _bool_op_to_ir(op: ast.boolop) -> Optional[ir.BinaryOperator]:
+    """Convert Python boolean operator to IR BinaryOperator."""
+    mapping = {
+        ast.And: ir.BinaryOperator.BINARY_OP_AND,
+        ast.Or: ir.BinaryOperator.BINARY_OP_OR,
+    }
+    return mapping.get(type(op))
+
+
+def _get_func_name(func: ast.expr) -> Optional[str]:
+    """Get function name from a Call's func attribute."""
+    if isinstance(func, ast.Name):
+        return func.id
+    elif isinstance(func, ast.Attribute):
+        # For method calls like obj.method, return full dotted name
+        parts = []
+        current = func
+        while isinstance(current, ast.Attribute):
+            parts.append(current.attr)
+            current = current.value
+        if isinstance(current, ast.Name):
+            parts.append(current.id)
+        name = ".".join(reversed(parts))
+        if name.startswith("self."):
+            return name[5:]
+        return name
+    return None
+
+
+def _is_self_method_call(node: ast.Call) -> bool:
+    """Return True if the call is a direct self.method(...) invocation."""
+    func = node.func
+    return (
+        isinstance(func, ast.Attribute)
+        and isinstance(func.value, ast.Name)
+        and func.value.id == "self"
+    )
+
+
+def _global_function_for_call(
+    func_name: str, node: ast.Call
+) -> Optional[ir.GlobalFunction.ValueType]:
+    """Return the GlobalFunction enum value for supported globals."""
+    if _is_self_method_call(node):
+        return None
+    return GLOBAL_FUNCTIONS.get(func_name)