write-operation

name	write-operation
description	Write, scaffold, or review an Artisan pipeline operation. Use this skill when the user asks to create a new operation, write a creator or curator, scaffold an operation class, or review an existing operation for correctness. Trigger on phrases like "write an operation", "create a creator", "new curator", "scaffold operation", or any request involving OperationDefinition subclasses.
argument-hint	[OperationClassName or description of what the operation should do]

Write an Artisan Operation

Write or scaffold an operation for $ARGUMENTS.

Before writing, read at least one example operation from src/artisan/operations/examples/ to match the established style. Also read the base class at src/artisan/operations/base/operation_definition.py if you need to confirm API details.

---

Decision: Creator or Curator?

Pick one. The framework detects the type by which method you override.

Type	Override	Use when	Runs on
Creator	execute()	Heavy computation, file I/O, external tools	Workers (ThreadPool / SLURM)
Curator	execute_curator()	Lightweight metadata: filter, merge, ingest, route	Local process

---

Creator Operation Template

Follow this structure exactly. Use the # ---------- Section ---------- comment style. Declare sections in this order: Metadata, Inputs, Outputs, Parameters, Resources, Execution, Lifecycle.

"""One-line module docstring describing what this operation does."""

from __future__ import annotations

from enum import StrEnum, auto
from typing import Any, ClassVar

from pydantic import BaseModel, Field

from artisan.operations.base.operation_definition import OperationDefinition
from artisan.schemas import ArtifactResult
from artisan.schemas.artifact.data import DataArtifact
from artisan.schemas.execution.execution_config import ExecutionConfig
from artisan.schemas.operation_config.resource_config import ResourceConfig
from artisan.schemas.specs.input_models import (
    ExecuteInput,
    PostprocessInput,
    PreprocessInput,
)
from artisan.schemas.specs.input_spec import InputSpec
from artisan.schemas.specs.output_spec import OutputSpec


class MyOperation(OperationDefinition):
    """One-line summary of what the operation does.

    Extended description if needed. Explain the algorithm or approach.
    """

    # ---------- Metadata ----------
    name = "my_operation"
    description = "One-line summary matching the docstring"

    # ---------- Inputs ----------
    class InputRole(StrEnum):
        DATASET = "dataset"

    inputs: ClassVar[dict[str, InputSpec]] = {
        InputRole.DATASET: InputSpec(
            artifact_type="data",
            required=True,
            description="Input CSV dataset",
        ),
    }

    # ---------- Outputs ----------
    class OutputRole(StrEnum):
        DATASET = "dataset"

    outputs: ClassVar[dict[str, OutputSpec]] = {
        OutputRole.DATASET: OutputSpec(
            artifact_type="data",
            description="Transformed dataset",
            infer_lineage_from={"inputs": ["dataset"]},
        ),
    }

    # ---------- Parameters ----------
    class Params(BaseModel):
        """Algorithm parameters for MyOperation."""

        threshold: float = Field(
            default=0.5,
            ge=0.0,
            le=1.0,
            description="Filtering threshold",
        )

    params: Params = Params()

    # ---------- Resources ----------
    resources: ResourceConfig = ResourceConfig(time_limit="00:30:00")

    # ---------- Execution ----------
    execution: ExecutionConfig = ExecutionConfig(job_name="my_operation")

    # ---------- Lifecycle ----------
    def preprocess(self, inputs: PreprocessInput) -> dict[str, Any]:
        """Extract materialized paths from input artifacts."""
        return {
            role: [a.materialized_path for a in artifacts]
            for role, artifacts in inputs.input_artifacts.items()
        }

    def execute(self, inputs: ExecuteInput) -> Any:
        """Core computation. Read from inputs, write to execute_dir."""
        ...

    def postprocess(self, inputs: PostprocessInput) -> ArtifactResult:
        """Build draft artifacts from execution outputs."""
        drafts = []
        for file_path in inputs.file_outputs:
            if file_path.suffix == ".csv":
                drafts.append(
                    DataArtifact.draft(
                        content=file_path.read_bytes(),
                        original_name=file_path.name,
                        step_number=inputs.step_number,
                    )
                )

        return ArtifactResult(
            success=True,
            artifacts={"dataset": drafts},
        )

---

Curator Operation Template

"""One-line module docstring."""

from __future__ import annotations

from enum import StrEnum, auto
from typing import TYPE_CHECKING, ClassVar

import polars as pl

from artisan.operations.base.operation_definition import OperationDefinition
from artisan.schemas.artifact.types import ArtifactTypes
from artisan.schemas.execution.curator_result import PassthroughResult
from artisan.schemas.specs.input_spec import InputSpec
from artisan.schemas.specs.output_spec import OutputSpec

if TYPE_CHECKING:
    from artisan.storage.core.artifact_store import ArtifactStore


class MyCurator(OperationDefinition):
    """One-line summary."""

    # ---------- Metadata ----------
    name = "my_curator"
    description = "One-line summary"

    # ---------- Inputs ----------
    class InputRole(StrEnum):
        stream = auto()

    inputs: ClassVar[dict[str, InputSpec]] = {
        InputRole.stream: InputSpec(
            artifact_type=ArtifactTypes.ANY,
            required=True,
            description="Artifacts to process",
        ),
    }

    # ---------- Outputs ----------
    class OutputRole(StrEnum):
        stream = auto()

    outputs: ClassVar[dict[str, OutputSpec]] = {
        OutputRole.stream: OutputSpec(
            artifact_type=ArtifactTypes.ANY,
            description="Processed artifacts",
        ),
    }

    # ---------- Lifecycle ----------
    def execute_curator(
        self,
        inputs: dict[str, pl.DataFrame],
        step_number: int,
        artifact_store: ArtifactStore,
    ) -> PassthroughResult:
        """Route artifact IDs without creating new artifacts."""
        ids = inputs["stream"]["artifact_id"].to_list()
        return PassthroughResult(
            success=True,
            passthrough={"stream": ids},
        )

---

Lineage Patterns

Every creator output must set infer_lineage_from. Curator outputs may use None.

Pattern	Value	When to use
Derived from input	{"inputs": ["role_name"]}	Output traces back to a named input role
Generative	{"inputs": []}	No parent artifacts (data generation)
Output-to-output	{"outputs": ["other_role"]}	Co-produced artifact (e.g. metrics alongside data)
Curator passthrough	None	Curator routing existing artifacts

Stem-matching rule: The framework infers 1:1 lineage by matching the original_name stem of output drafts against input artifact names. For transforms, preserve the input filename stem in the output name.

---

InputSpec Fields

Field	Type	Default	Effect
artifact_type	str	"any"	Type constraint on accepted artifacts
required	bool	True	Pipeline fails if input is missing
materialize	bool	True	True: write to disk (file path). False: in-memory access only
materialize_as	str | None	None	Target format for materialization (e.g. ".pdb"). Requires materialize=True
hydrate	bool	True	True: load content. False: ID-only mode
with_associated	tuple[str, ...]	()	Auto-resolve related artifacts via provenance

OutputSpec Fields

Field	Type	Effect
artifact_type	str	Type of artifact produced
description	str	Human-readable description
required	bool	Whether the output must be non-empty
infer_lineage_from	dict | None	Lineage declaration (see patterns above)

infer_lineage_from constraints: Empty dict {} is invalid (raises ValidationError). Combined {"inputs": [...], "outputs": [...]} is not supported — use separate output roles instead.

ExecuteInput Fields

Field	Type	Default	Effect
execute_dir	Path	—	Directory for writing output files
inputs	dict[str, Any]	{}	Prepared inputs from preprocess()
log_path	Path | None	None	Path for external tool stdout/stderr capture
metadata	dict[str, Any]	{}	Extensibility escape hatch from the engine

PostprocessInput Fields

Field	Type	Default	Effect
step_number	int	—	Current pipeline step number (required for draft())
postprocess_dir	Path	—	Directory for postprocess artifacts (rarely needed)
file_outputs	list[Path]	[]	All files in execute_dir after execute completes
memory_outputs	Any	None	Whatever execute() returned
input_artifacts	dict[str, list[Artifact]]	{}	Full input context with metadata for output naming and lineage
metadata	dict[str, Any]	{}	Extensibility escape hatch from the engine

PostprocessInput also provides associated_artifacts(artifact, type_str) and grouped() — the same methods available on PreprocessInput. Operations that need input context in postprocess (e.g., propagating annotations) use inputs.input_artifacts and inputs.associated_artifacts() directly.

---

Variant: Generative Creator (No Inputs)

• Omit InputRole
• Set inputs: ClassVar[dict] = {}
• Set infer_lineage_from={"inputs": []} on all outputs
• Implement only execute() and postprocess() (no preprocess())

See src/artisan/operations/examples/data_generator.py.

Variant: Multi-input with group_by

• Define multiple roles in InputRole and inputs
• Set group_by: ClassVar[GroupByStrategy | None] = GroupByStrategy.LINEAGE (or ZIP or CROSS_PRODUCT)
• In preprocess, iterate paired groups via inputs.grouped()

def preprocess(self, inputs: PreprocessInput) -> dict[str, Any]:
    prepared = []
    for group in inputs.grouped():
        prepared.append({
            "dataset": str(group["dataset"].materialized_path),
            "config": str(group["config"].materialized_path),
        })
    return {"items": prepared}

See src/artisan/operations/examples/data_transformer_script.py.

Variant: External Tool via ToolSpec

• Set tool: ToolSpec = ToolSpec(executable=SCRIPT_PATH, interpreter="python")
• Configure environments: Environments = Environments(local=..., docker=...)
• In execute, call run_command(env, [*self.tool.parts(), *args])
• Import from artisan.utils.external_tools import format_args, run_command

See src/artisan/operations/examples/data_transformer_script.py.

Variant: Config Artifacts with $artifact References

• Produce ExecutionConfigArtifact drafts with {"$artifact": artifact_id} placeholders in the content dict
• The framework resolves $artifact references to materialized paths at execution time
• Set materialize=False on the input to access artifact IDs without writing files to disk

See src/artisan/operations/examples/data_transformer_config.py.

Variant: Output-to-output Lineage

• Use infer_lineage_from={"outputs": ["other_role"]} when one output derives from another co-produced output (e.g. metrics computed alongside data)
• The primary output uses {"inputs": []} or {"inputs": ["role"]}
• The derived output points to the primary via {"outputs": ["primary_role"]}

See src/artisan/operations/examples/data_generator_with_metrics.py.

---

Validation Rules

The framework validates at class definition time (import). These cause TypeError immediately:

• Must override either execute() or execute_curator() (not neither, not both)
• Creator outputs must set infer_lineage_from (cannot be None)
• Creator operations with inputs must implement preprocess()
• Must define OutputRole(StrEnum) with values matching outputs keys exactly
• Must define InputRole(StrEnum) with values matching inputs keys exactly (unless inputs is empty or runtime_defined_inputs=True)

---

Artifact Draft Methods

Use the appropriate draft() class method in postprocess:

# File-based data (CSV, binary, etc.)
DataArtifact.draft(content=bytes, original_name=str, step_number=int)

# Key-value metrics (JSON-serializable dict)
MetricArtifact.draft(content=dict, original_name=str, step_number=int)

# Execution configs with $artifact references
ExecutionConfigArtifact.draft(content=dict, original_name=str, step_number=int)

The original_name is critical: the framework uses filename stems for lineage matching. For 1:1 transforms, preserve the input's filename stem.

---

Testing Patterns

Unit test a creator

def test_my_operation(tmp_path):
    op = MyOperation(params=MyOperation.Params(threshold=0.8))

    # Prepare input files
    input_csv = tmp_path / "input.csv"
    input_csv.write_text("id,value\n1,0.9\n2,0.3\n")

    execute_dir = tmp_path / "execute"
    execute_input = ExecuteInput(
        execute_dir=execute_dir,
        inputs={"dataset": [str(input_csv)]},
    )
    memory_outputs = op.execute(execute_input)

    post_input = PostprocessInput(
        step_number=0,
        postprocess_dir=tmp_path / "post",
        file_outputs=list(execute_dir.iterdir()),
        memory_outputs=memory_outputs,
    )
    result = op.postprocess(post_input)
    assert result.success
    assert "dataset" in result.artifacts

Unit test a curator

def test_my_curator():
    op = MyCurator()
    result = op.execute_curator(
        inputs={"stream": pl.DataFrame({"artifact_id": ["abc123", "def456"]})},
        step_number=1,
        artifact_store=Mock(),
    )
    assert result.success
    assert len(result.passthrough["stream"]) == 2

---

Style Rules

Follow these conventions from the existing examples:

• Module docstring: One line, describes what the operation does
• Class docstring: Summary line + optional extended description. Do not manually write Input/Output Roles sections (auto-generated by the framework)
• Section comments: Use # ---------- Section ---------- with exactly 10 dashes on each side
• Section order: Metadata, Inputs, Outputs, Parameters, Resources, Execution, Lifecycle (omit sections that use defaults)
• Lifecycle docstrings: One-line imperative summary (e.g. "Extract materialized paths from input artifacts.")
• name value: snake_case matching the class name's snake_case form
• Imports: Group stdlib, then pydantic, then artisan. Use from __future__ import annotations
• Params class: Nest inside the operation class. Use Field() with default, constraints (ge, le), and description for every parameter
• No bare constants: Put algorithm-specific values in Params, not as module-level constants
• execute() is a black box: It reads files and writes files. No framework imports, no Artifact objects, no ArtifactStore access
• preprocess() bridges in: Converts Artifact objects to plain paths/dicts
• postprocess() bridges out: Converts files/memory_outputs to draft Artifacts
• Return metadata: Include operation name and key params in ArtifactResult.metadata