Perceptic Workflow SDK

Python SDK for Perceptic Workflow definitions. This package provides cross-language Temporal workflow types that are compatible with the Java workflow definitions in perceptic-core-client.

Installation

uv add perceptic-workflow-sdk

Usage

Runtime Input Contract

Workflows started through Perceptic Core receive plain JSON input data.

• The first @workflow.run argument should be a JSON-compatible structure (dict[str, Any] by default).
• Perceptic Core does not wrap inputs in an SDK envelope type.
• Consumers are responsible for interpreting and validating the JSON shape they expect.

Workflow Events

from perceptic_workflow import (
    WorkflowEvent,
    InfoEvent,
    CheckpointEvent,
    UserInputEvent,
    UserInputRequestEvent,
    WorkflowCheckpointStatus,
)

# Create events
info_event = InfoEvent(
    event_id=1,
    type="progress",
    payload={"step": "processing"},
    timestamp=datetime.now()
)

Implementing Workflows

The SDK provides a mixin class for implementing Perceptic-compatible workflows:

from typing import Any
from pydantic import BaseModel
from temporalio import workflow
from perceptic_workflow import PercepticWorkflow, WorkflowEvent


class ResearchInput(BaseModel):
    query: str
    context: str

@workflow.defn
class MyWorkflow(PercepticWorkflow):
    def __init__(self):
        self._events: list[WorkflowEvent] = []
        self._paused = False

    @workflow.run
    async def run(self, inputs: dict[str, Any]) -> dict:
        request = ResearchInput.model_validate(inputs)

        return {
            "query": request.query,
            "context": request.context,
        }

    @workflow.query(name=PercepticWorkflow.QUERY_GET_STATUS)
    def get_status(self) -> WorkflowStatus:
        return WorkflowStatus(status="running", description="Waiting for input" if self._paused else "Running")

    @workflow.query(name=PercepticWorkflow.QUERY_GET_EVENTS)
    def get_events(self, after_event_id: int | None = None) -> dict[str, Any]:
        if after_event_id is None:
            return {"events": self._events}
        return {"events": [e for e in self._events if e["eventId"] > after_event_id]}

Decorator Validation

PercepticWorkflow no longer performs runtime validation of Temporal handler decorators during class creation. This avoids import-time failures for abstract workflow bases and mixin-based inheritance patterns.

Handler contract enforcement is expected to happen through linting and type-check tooling in your development pipeline:

uv run ty check
uv run workflow-typecheck

workflow-typecheck implementation notes

workflow-typecheck is implemented by src/perceptic_workflow/_checks/workflow_decorator_check.py and performs a static AST scan of Python classes that inherit from PercepticWorkflow (or from a class that already does). For the known handler methods, it enforces:

• the method has a Temporal handler decorator (@workflow.query or @workflow.update)
• the decorator kind matches the expected contract
• the name= argument matches the PercepticWorkflow handler constant

Current diagnostics:

• PWF001: expected Temporal decorator is missing
• PWF002: wrong decorator kind (query vs update)
• PWF003: wrong or missing handler name=...

How to extend the check:

• Add or update required handlers in REQUIRED_HANDLERS when a new Perceptic handler method is added or an existing one changes.
• Add base class names to MIXIN_COMPATIBLE_BASES if additional workflow base classes should be recognized as Perceptic-compatible roots.
• Extend _extract_name_kwarg / _is_workflow_decorator_call if we support additional decorator call shapes or name expressions.

When this check must be updated:

• A PercepticWorkflow handler method is renamed, added, removed, or changes from query to update (or vice versa).
• Any PercepticWorkflow.<HANDLER_CONSTANT> used in handler decorators is renamed or replaced.
• We introduce a new inheritance pattern for workflow mixins or a new canonical workflow base class.
• We change how Temporal decorators are authored (for example, different decorator access paths or non-current name= expression style).

How to edit safely:

1. Update the contract mappings first (REQUIRED_HANDLERS, MIXIN_COMPATIBLE_BASES).
2. Keep diagnostics specific and actionable (PWF001-PWF003 style).
3. Add or update tests in tests/test_workflow_typecheck_regression.py for the new/changed behavior.
4. Validate with:
   • uv run workflow-typecheck
   • uv run pytest tests/test_workflow_typecheck_regression.py
   • uv run ty check

Troubleshooting

• ValidationError during workflow activation usually means your @workflow.run signature expects the wrong input shape.
• Use a JSON-compatible first workflow argument (for example dict[str, Any]), then validate/coerce into your domain model inside run.
• For run-start requests, send inputs only; Perceptic Core forwards that JSON payload directly.

Compatibility

This package is designed to be compatible with:

• perceptic-core-client (Java)
• perceptic-core-workflow-runtime (Java)

The workflow definitions and event types are generated from the same JSON Schema sources to ensure cross-language compatibility.