objectstate

Generic lazy dataclass configuration framework with dual-axis inheritance

Features

• Lazy Dataclass Factory: Dynamically create dataclasses with lazy field resolution
• Dual-Axis Inheritance:
  • X-Axis: Context hierarchy (Step → Pipeline → Global)
  • Y-Axis: Sibling inheritance within same context
• Contextvars-Based: Uses Python's contextvars for clean context management
• UI Integration: Placeholder text generation for configuration forms
• Thread-Safe: Thread-local global configuration storage
• 100% Generic: No application-specific dependencies
• Small dependency surface: Uses python-introspect for callable and dataclass analysis
• Custom Dataclass Rebuild Hook: For init=False dataclasses during lazy serialization (__objectstate_rebuild__)

Quick Start

Simple Usage (Manual Factory)

from dataclasses import dataclass
from objectstate import LazyDataclassFactory, config_context, set_base_config_type

# Define your base configuration
@dataclass
class MyConfig:
    output_dir: str = "/tmp"
    num_workers: int = 4
    debug: bool = False

# Initialize the framework with your base config type
set_base_config_type(MyConfig)

# Create lazy version manually
LazyMyConfig = LazyDataclassFactory.make_lazy_simple(MyConfig)

# Use with context
concrete_config = MyConfig(output_dir="/data", num_workers=8)

with config_context(concrete_config):
    lazy_cfg = LazyMyConfig()
    print(lazy_cfg.output_dir)  # "/data" (resolved from context)
    print(lazy_cfg.num_workers)  # 8 (resolved from context)
    print(lazy_cfg.debug)        # False (inherited from defaults)

Project a concrete config into its generated lazy type with from_config. This is the supported boundary when an application has already parsed or loaded concrete configuration:

concrete = MyConfig(output_dir="/data", num_workers=8)
lazy_cfg = LazyMyConfig.from_config(concrete)

# Keep only values that differ from an inherited concrete config.
inherited = MyConfig(output_dir="/data", num_workers=4)
overrides = LazyMyConfig.from_config(concrete, inherited=inherited)

from_config accepts concrete dataclass instances. Passing another lazy instance is an error because it would make the projection depend on ambient resolution state.

Setting Up Global Config Context (For Advanced Usage)

When using the decorator pattern with auto_create_decorator, you need to establish the global configuration context for lazy resolution:

from objectstate import ensure_global_config_context

# After creating your global config instance
global_config = GlobalPipelineConfig(
    num_workers=8,
    # ... other fields
)

# REQUIRED: Establish global config context for lazy resolution
ensure_global_config_context(GlobalPipelineConfig, global_config)

# Now lazy configs can resolve from the global context

Key differences:

• set_base_config_type(MyConfig): Sets the type (class) for the framework
• ensure_global_config_context(GlobalConfig, instance): Sets the instance (concrete values) for resolution
• Call ensure_global_config_context() at application startup (GUI) or before pipeline execution

Installation

pip install objectstate

ObjectState Registry

ObjectState separates mutable working state from saved baseline, enabling dirty tracking and undo/redo:

from objectstate import ObjectState, ObjectStateRegistry

# Register an object (e.g., when added to pipeline)
state = ObjectState(my_step_config, scope_id="/pipeline::step_0")
ObjectStateRegistry.register(state)

# Query the registry
state = ObjectStateRegistry.get_by_scope("/pipeline::step_0")
all_states = ObjectStateRegistry.get_all()

# Update a parameter (marks field as dirty)
state.update_parameter("output_dir", "/new/path")

# Check dirty state
if state.dirty_fields:
    print(f"Unsaved changes: {state.dirty_fields}")

# Save changes (updates baseline)
state.mark_saved()

# Or restore to saved baseline
state.restore_saved()

# Unregister when removed
ObjectStateRegistry.unregister(state)

Undo/Redo and Time Travel

Git-like DAG history with branching timelines:

from objectstate import ObjectStateRegistry

# Time travel (automatically recorded on parameter changes)
ObjectStateRegistry.time_travel_back()   # Go one step back in history
ObjectStateRegistry.time_travel_forward()  # Go one step forward in history

# Note: ObjectState uses time-travel semantics (like Git), not traditional undo/redo.
# You can navigate to any point in history and make new changes (creating branches).
# This is more powerful than undo/redo for complex workflows.

# Batch multiple changes into one snapshot
with ObjectStateRegistry.atomic("add item"):
    ObjectStateRegistry.register(item_state)
    parent_state.update_parameter("items", new_items)

# Time travel to specific point
history = ObjectStateRegistry.get_branch_history()
ObjectStateRegistry.time_travel_to_snapshot(history[5].id)
ObjectStateRegistry.time_travel_to_head()  # Return to latest

# Branching timelines
ObjectStateRegistry.create_branch("experiment", description="Testing new approach")
ObjectStateRegistry.switch_branch("main")
branches = ObjectStateRegistry.list_branches()

# Persist history
history_dict = ObjectStateRegistry.export_history_to_dict()
ObjectStateRegistry.import_history_from_dict(history_dict)

# Or save to file
ObjectStateRegistry.save_history_to_file("history.objectstate")
ObjectStateRegistry.load_history_from_file("history.objectstate")

Automatic Lazy Config Generation with Decorators

For more complex applications with multiple config types, use the auto_create_decorator pattern to automatically generate lazy versions and field injection decorators:

from dataclasses import dataclass
from objectstate import auto_create_decorator, config_context

# Step 1: Create a global config class with "Global" prefix and apply auto_create_decorator
@auto_create_decorator
@dataclass
class GlobalPipelineConfig:
    base_output_dir: str = "/tmp"
    verbose: bool = False

# This automatically creates:
# - A decorator named `global_pipeline_config` (snake_case of class name)
#   that you can use to decorate other config classes
# - A lazy class `PipelineConfig` (removes "Global" prefix) for lazy resolution

# Step 2: Use the generated decorator on other config classes
@global_pipeline_config  # Automatically creates LazyStepConfig
@dataclass
class StepConfig:
    step_name: str = "default_step"
    iterations: int = 100

@global_pipeline_config  # Automatically creates LazyDatabaseConfig  
@dataclass
class DatabaseConfig:
    host: str = "localhost"
    port: int = 5432

# The decorator automatically:
# - Creates lazy versions: LazyStepConfig, LazyDatabaseConfig
# - Registers them for potential field injection into GlobalPipelineConfig
# - Makes them available in your module namespace

Key Benefits:

• Auto-generated lazy classes: Each decorated config automatically gets a lazy version
• Simplified imports: Lazy classes are automatically added to your module
• Decorator factory: auto_create_decorator generates a decorator specific to your global config
• Type-safe: All generated classes are proper dataclasses with full IDE support

Field Injection Behavior

When you use the generated decorator (e.g., @global_pipeline_config), the decorated class is automatically injected as a field into the global config class:

from dataclasses import dataclass
from objectstate import auto_create_decorator

# Create global config with auto_create_decorator
@auto_create_decorator
@dataclass
class GlobalPipelineConfig:
    num_workers: int = 1

# This creates:
# - A decorator named `global_pipeline_config`
# - A lazy class named `PipelineConfig`

# Use the decorator on a new config class
@global_pipeline_config
@dataclass
class WellFilterConfig:
    well_filter: str = None
    mode: str = "include"

# After module loading, GlobalPipelineConfig automatically has:
# - well_filter_config: WellFilterConfig = WellFilterConfig()
# And LazyWellFilterConfig is auto-created

How it works:

• Decorated classes are injected as fields into GlobalPipelineConfig
• Field name is snake_case of class name (e.g., WellFilterConfig → well_filter_config)
• Lazy version is automatically created (e.g., LazyWellFilterConfig)
• Injection happens at end of module loading via _inject_all_pending_fields()

This enables a clean, modular configuration structure where each component's config is automatically part of the global configuration.

Decorator Parameters

The generated decorator (e.g., @global_pipeline_config) supports optional parameters:

inherit_as_none (Default: True)

Sets all inherited fields from parent classes to None by default, enabling proper lazy resolution:

@dataclass
class BaseConfig:
    timeout: int = 30
    retries: int = 3

@global_pipeline_config(inherit_as_none=True)  # Default behavior
@dataclass
class ServiceConfig(BaseConfig):
    service_name: str = "my-service"
    # timeout and retries automatically set to None for lazy inheritance

# This allows ServiceConfig to inherit timeout/retries from context
# rather than using the base class defaults

Why this matters:

• Enables polymorphic access without type-specific attribute names
• Critical for dual-axis inheritance with multiple inheritance
• Uses InheritAsNoneMeta metaclass internally

ui_hidden (Default: False)

Hides configs from UI while still applying decorator behavior and keeping them in the resolution context:

@global_pipeline_config(ui_hidden=True)
@dataclass
class InternalConfig:
    internal_setting: str = "hidden"
    # This config won't appear in UI but is still available for inheritance

Use cases:

• Intermediate configs that are only inherited by other configs
• Internal implementation details not meant for user configuration
• Base classes that should never be directly instantiated in UI

Nested Dataclass Lazification

When creating a lazy dataclass, nested dataclass fields are automatically converted to their lazy versions:

from dataclasses import dataclass
from objectstate import LazyDataclassFactory

@dataclass
class DatabaseConfig:
    host: str = "localhost"
    port: int = 5432

@dataclass
class AppConfig:
    db_config: DatabaseConfig = DatabaseConfig()
    app_name: str = "MyApp"

# Create lazy version - nested configs are automatically lazified
LazyAppConfig = LazyDataclassFactory.make_lazy_simple(AppConfig)

# The db_config field is automatically converted to LazyDatabaseConfig
# You don't need to manually create LazyDatabaseConfig first!

Benefits:

• No need to manually create lazy versions of nested configs
• Preserves field metadata (e.g., ui_hidden flag)
• Creates default factories for Optional dataclass fields
• Uses register_lazy_type_mapping() internally

Why objectstate?

Before (Manual parameter passing):

def process_step(data, output_dir, num_workers, debug, *more_options):
    # Pass 20+ parameters through every function
    result = sub_process(data, output_dir, num_workers, debug, *more_options)
    return result

def sub_process(data, output_dir, num_workers, debug, *more_options):
    # Repeat parameter declarations everywhere
    ...

After (objectstate):

@dataclass
class StepConfig:
    output_dir: str = None
    num_workers: int = None
    debug: bool = None

def process_step(data, config: LazyStepConfig):
    # Config fields resolve automatically from context
    print(config.output_dir)  # Resolved from context hierarchy
    result = sub_process(data, config)
    return result

Advanced Features

Dual-Axis Inheritance

# X-Axis: Context hierarchy
with config_context(global_config):
    with config_context(pipeline_config):
        with config_context(step_config):
            # Resolves: step → pipeline → global → defaults
            value = objectstate.some_field

# Y-Axis: Sibling inheritance (MRO-based)
@dataclass
class BaseConfig:
    field_a: str = "base"

@dataclass
class SpecializedConfig(BaseConfig):
    field_b: str = "specialized"

# SpecializedConfig inherits field_a from BaseConfig

Accessing Resolved Values

from objectstate import ObjectState, ObjectStateRegistry

# ObjectState stores both saved baseline and live (edited) values
state = ObjectStateRegistry.get_by_scope("/pipeline::step_0")

# Access resolved value (from _live_resolved cache)
output_dir = state.get_resolved_value("output_dir")

# Check if field is dirty (live != saved)
is_dirty = "output_dir" in state.dirty_fields

# Get provenance (where did this value come from?)
source_scope, source_type = state.get_resolved_provenance("output_dir")
# Returns: ("/pipeline", GlobalPipelineConfig) if inherited from pipeline

Architecture

Dual-Axis Resolution

The framework uses pure MRO-based dual-axis resolution:

X-Axis (Context Hierarchy):

Step Context → Pipeline Context → Global Context → Static Defaults

Y-Axis (MRO Traversal):

Most specific class → Least specific class (following Python's MRO)

How it works:

1. Context hierarchy is flattened into a single available_configs dict
2. For each field resolution, traverse the requesting object's MRO from most to least specific
3. For each MRO class, check if there's a config instance in available_configs with a concrete (non-None) value
4. Return the first concrete value found

Parametric Axes Prototype (PEP Draft)

The parametric_axes module demonstrates extending Python's type system with arbitrary semantic axes beyond (B, S):

from objectstate.parametric_axes import AxesBase

class Step(AxesBase):
    pass

class MyStep(Step, axes={"scope": "/pipeline/step_0", "registry": "handlers"}):
    pass

MyStep.__axes__  # {'scope': '/pipeline/step_0', 'registry': 'handlers'}
MyStep.__scope__  # '/pipeline/step_0' (convenience attribute)

This works TODAY via __init_subclass__ (PEP 487) - no grammar changes required.

Three usage patterns:

• AxesBase inheritance: class Foo(Base, axes={...}) (preferred)
• Factory function: axes_type("Foo", (Base,), {}, scope="...", registry="...")
• Decorator: @with_axes(scope="...", registry="...") (when base can't be modified)

Try It Out

Run the tests to see the prototype in action:

# Run all parametric axes tests
python -m pytest tests/test_parametric_axes.py -v

# Interactive exploration
python -c "
from objectstate.parametric_axes import AxesBase, axes_type, with_axes, get_axes

# Pattern 1: Class statement syntax (preferred)
class Step(AxesBase):
    pass

class ProcessingStep(Step, axes={'scope': '/pipeline/step_0', 'registry': 'processing'}):
    def process(self):
        return 'processed'

print(f'ProcessingStep.__axes__ = {dict(ProcessingStep.__axes__)}')
print(f'ProcessingStep.__scope__ = {ProcessingStep.__scope__}')

# Pattern 2: Factory function (mimics extended type())
Handler = axes_type('Handler', (), {}, format='imagexpress', version=2)
print(f'Handler.__axes__ = {Handler.__axes__}')

# Pattern 3: Decorator (when base can't be modified)
@with_axes(scope='/decorated', priority=10)
class DecoratedStep:
    pass
print(f'DecoratedStep.__axes__ = {DecoratedStep.__axes__}')

# MRO-based axis resolution with multiple inheritance
class A(AxesBase, axes={'x': 1, 'from_a': True}):
    pass
class B(AxesBase, axes={'x': 2, 'from_b': True}):
    pass
class C(A, B):  # x=1 from A (leftmost in MRO)
    pass
print(f'C.__axes__ = {dict(C.__axes__)}')  # x=1, from_a=True, from_b=True
"

See src/objectstate/parametric_axes.py for full implementation and docstrings.

Reified Generics Prototype

The reified_generics module provides runtime-accessible type parameters for generics:

from objectstate.reified_generics import List, Dict

# Type parameters are preserved at runtime
IntList = List[int]
StrDict = Dict[str, float]

# Introspection works
IntList.__args__      # (int,)
IntList.__origin__    # list
StrDict.__args__      # (str, float)

# isinstance checks work with reified types
my_list = IntList([1, 2, 3])
isinstance(my_list, IntList)  # True
isinstance(my_list, List[str])  # False - different type parameter!

# Type caching ensures identity
List[int] is List[int]  # True - same object

Try It Out

# Run reified generics tests
python -m pytest tests/test_reified_generics.py -v

# Interactive exploration
python -c "
from objectstate.reified_generics import List, Dict, Set, Optional

# Create reified types
IntList = List[int]
StrIntDict = Dict[str, int]

print(f'IntList.__args__ = {IntList.__args__}')
print(f'IntList.__origin__ = {IntList.__origin__}')
print(f'StrIntDict.__args__ = {StrIntDict.__args__}')

# Type identity (caching)
print(f'List[int] is List[int]: {List[int] is List[int]}')
print(f'List[int] is List[str]: {List[int] is List[str]}')

# Create instances
my_list = IntList([1, 2, 3])
print(f'isinstance(my_list, IntList): {isinstance(my_list, IntList)}')
print(f'isinstance(my_list, List[str]): {isinstance(my_list, List[str])}')
"

See src/objectstate/reified_generics.py for full implementation.

Documentation

Full documentation available at objectstate.readthedocs.io

Requirements

• Python 3.11+
• python-introspect is the only runtime dependency

License

MIT License - see LICENSE file for details

Contributing

Contributions welcome! Please see CONTRIBUTING.md for guidelines.

Credits

Developed by Tristan Simas as part of the OpenHCS project.