Synapse Action Development

Synapse SDK provides two patterns for plugin actions: function-based (simple, stateless) and class-based (complex, stateful).

Quick Start: Function-Based Action

from pydantic import BaseModel
from synapse_sdk.plugins.decorators import action
from synapse_sdk.plugins.context import RuntimeContext

class TrainParams(BaseModel):
    epochs: int = 10
    learning_rate: float = 0.001

@action(name='train', description='Train a model', params=TrainParams)
def train(params: TrainParams, ctx: RuntimeContext) -> dict:
    for epoch in range(params.epochs):
        ctx.set_progress(epoch + 1, params.epochs)
    return {'status': 'completed'}

Quick Start: Class-Based Action

from pydantic import BaseModel
from synapse_sdk.plugins.action import BaseAction

class InferParams(BaseModel):
    model_path: str
    threshold: float = 0.5

class InferAction(BaseAction[InferParams]):
    action_name = 'inference'

    def execute(self) -> dict:
        self.set_progress(0, 100)
        # Implementation here
        return {'predictions': []}

When to Use Each Pattern

| Criteria | Function-Based | Class-Based | |----------|----------------|-------------| | Complexity | Simple, single-purpose | Complex, multi-step | | State | Stateless | Can use helper methods | | Semantic types | Limited | Full support |

Recommendation: Start with function-based. Use class-based when needing helper methods or semantic type declarations.

@action Decorator Parameters

| Parameter | Required | Description | |-----------|----------|-------------| | name | No | Action name (defaults to function name) | | description | No | Human-readable description | | params | No | Pydantic model for parameter validation | | result | No | Pydantic model for result validation | | category | No | PluginCategory for grouping |

Category Parameter Examples

from synapse_sdk.plugins.decorators import action
from synapse_sdk.plugins.constants import PluginCategory

# Training action
@action(
    name='train',
    category=PluginCategory.NEURAL_NET,
    description='Train object detection model'
)
def train(params, ctx):
    ...

# Export action
@action(
    name='export_coco',
    category=PluginCategory.EXPORT,
    description='Export to COCO format'
)
def export_coco(params, ctx):
    ...

# Smart tool (AI-assisted annotation)
@action(
    name='auto_segment',
    category=PluginCategory.SMART_TOOL,
    description='Auto-segmentation tool'
)
def auto_segment(params, ctx):
    ...

# Pre-annotation
@action(
    name='pre_label',
    category=PluginCategory.PRE_ANNOTATION,
    description='Pre-label with model predictions'
)
def pre_label(params, ctx):
    ...

Available Categories: NEURAL_NET, EXPORT, UPLOAD, SMART_TOOL, PRE_ANNOTATION, POST_ANNOTATION, DATA_VALIDATION, CUSTOM

BaseAction Class Attributes

| Attribute | Description | |-----------|-------------| | action_name | Action name for invocation | | category | PluginCategory | | input_type | Semantic input type for pipelines | | output_type | Semantic output type for pipelines | | params_model | Auto-extracted from generic | | result_model | Optional result schema |

Available Methods in BaseAction

• self.params - Validated parameters
• self.ctx - RuntimeContext
• self.logger - Logger shortcut
• self.set_progress(current, total, category) - Progress tracking
• self.set_metrics(value, category) - Metrics recording
• self.log(event, data, file) - Event logging

Additional Resources

For detailed patterns and advanced techniques:

• references/patterns.md - Parameter validation, semantic types
• references/categories.md - Available PluginCategory values
• references/types-hierarchy.md - Semantic type system (DataType, Dataset, Model)
• references/pipelines.md - ActionPipeline for chaining actions