write-composite

name	write-composite
description	Write or scaffold an Artisan composite operation. Use this skill when the user asks to create a composite, write a composite operation, scaffold a composite, compose operations, or any request involving CompositeDefinition subclasses. Trigger on phrases like "write a composite", "create a composite operation", "scaffold a composite", "compose operations", or any request to combine multiple operations into a reusable unit.
argument-hint	[CompositeClassName or description of what the composite should do]

Write an Artisan Composite

Write or scaffold a composite operation for $ARGUMENTS.

Before writing, read the example composites in integration tests (tests/integration/test_composite_collapsed.py, tests/integration/test_composite_expanded.py) and the base class at src/artisan/composites/base/composite_definition.py to match established patterns.

---

When to Use a Composite

• Reusable multi-operation sequence — a fixed chain of operations that appears in multiple pipelines
• Tightly coupled ops — operations that logically belong together and should run as a single pipeline step (collapsed) or be expandable for debugging
• Dual-mode execution — same definition works with pipeline.run() (collapsed, single step) and pipeline.expand() (expanded, one step per internal op)

If you only need a single operation, use write-operation instead.

---

Composite Template

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

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

from __future__ import annotations

from enum import StrEnum
from typing import ClassVar

from artisan.composites import CompositeContext, CompositeDefinition
from artisan.schemas.specs.input_spec import InputSpec
from artisan.schemas.specs.output_spec import OutputSpec

# Import operations used in compose()
from artisan.operations.examples.data_transformer import DataTransformer
from artisan.operations.examples.metric_calculator import MetricCalculator


class TransformAndScore(CompositeDefinition):
    """Transform data then compute metrics.

    Runs DataTransformer followed by MetricCalculator, exposing only the
    final metrics as output.
    """

    # ---------- Metadata ----------
    name = "transform_and_score"
    description = "Transform data then compute metrics"

    # ---------- Inputs ----------
    class InputRole(StrEnum):
        DATA = "data"

    inputs: ClassVar[dict[str, InputSpec]] = {
        InputRole.DATA: InputSpec(
            artifact_type="data",
            required=True,
            description="Input dataset to transform and score",
        ),
    }

    # ---------- Outputs ----------
    class OutputRole(StrEnum):
        METRICS = "metrics"

    outputs: ClassVar[dict[str, OutputSpec]] = {
        OutputRole.METRICS: OutputSpec(
            artifact_type="metric",
            description="Computed metrics from the transformed data",
        ),
    }

    # ---------- Compose ----------
    def compose(self, ctx: CompositeContext) -> None:
        """Define the internal operation graph."""
        transformed = ctx.run(
            DataTransformer,
            inputs={"dataset": ctx.input("data")},
            params={"scale_factor": 2.0},
        )
        metrics = ctx.run(
            MetricCalculator,
            inputs={"dataset": transformed.output("dataset")},
        )
        ctx.output("metrics", metrics.output("metrics"))

---

compose() Method

The compose() method defines the internal operation graph using three primitives on the CompositeContext:

Method	Purpose
ctx.input(role)	Get a CompositeRef for an external input by role name
ctx.run(operation, inputs=..., params=..., ...)	Run an internal operation, returns CompositeStepHandle
ctx.output(role, ref)	Map an internal output to an external output role

ctx.run() returns a CompositeStepHandle whose .output(role) returns a CompositeRef for wiring to downstream internal operations or to ctx.output().

---

Per-Operation Overrides in compose()

ctx.run() accepts optional overrides that are forwarded to the operation:

ctx.run(
    DataTransformer,
    inputs={"dataset": ctx.input("data")},
    params={"scale_factor": 2.0},
    resources={"cpus": 4, "memory_gb": 16},
    execution={"artifacts_per_unit": 10},
    backend="slurm",
    environment="my_container",
    tool={"executable": "/path/to/tool"},
)

In collapsed mode, resources/backend/environment/tool are ignored (the composite runs as a single unit). In expanded mode, they are forwarded to each pipeline step.

---

Nesting Composites

Composites can contain other composites via ctx.run():

class OuterComposite(CompositeDefinition):
    name = "outer"
    description = "Runs an inner composite then scores"

    class OutputRole(StrEnum):
        METRICS = "metrics"

    outputs: ClassVar[dict[str, OutputSpec]] = {
        "metrics": OutputSpec(artifact_type="metric"),
    }

    def compose(self, ctx: CompositeContext) -> None:
        inner = ctx.run(InnerComposite)
        metrics = ctx.run(MetricCalculator,
            inputs={"dataset": inner.output("dataset")})
        ctx.output("metrics", metrics.output("metrics"))

---

Intermediates Modes

Controls what happens to artifacts from internal operations (collapsed mode only):

Mode	Behavior
"discard" (default)	Only final outputs are committed; intermediates are discarded
"persist"	Intermediates committed to Delta but not visible as step boundaries
"expose"	Intermediates committed to Delta with full step boundaries

Set via pipeline.run(MyComposite, intermediates="persist", ...). Not applicable to pipeline.expand() where all steps are naturally visible.

---

Differences from OperationDefinition

Composites are not operations. Key differences:

Feature	OperationDefinition	CompositeDefinition
Core method	execute() or execute_curator()	compose()
Lifecycle hooks	preprocess(), postprocess()	None
tool / environments	Supported	Not supported (set per-op in compose)
infer_lineage_from on outputs	Required for creators	Not supported
runtime_defined_inputs	Supported	Not supported
group_by	Supported	Not supported
Registry	OperationDefinition._registry	CompositeDefinition._registry

---

Validation Rules

The framework validates at class definition time (__pydantic_init_subclass__):

• compose() must be overridden (raises TypeError otherwise)
• OutputRole(StrEnum) values must match outputs keys exactly
• InputRole(StrEnum) values must match inputs keys exactly (if inputs are defined)
• Registered automatically in CompositeDefinition._registry by name

---

Running Composites

Two execution modes — see write-pipeline for full API details:

# Collapsed: single pipeline step, in-memory passing
pipeline.run(TransformAndScore, name="ts",
    inputs={"data": output("generate", "datasets")},
    intermediates="discard")

# Expanded: each internal op becomes its own pipeline step
expanded = pipeline.expand(TransformAndScore, name="ts",
    inputs={"data": output("generate", "datasets")})
pipeline.run(NextOp, inputs={"data": expanded.output("metrics")})

---

Testing Patterns

Integration test (collapsed)

def test_composite_collapsed(tmp_path):
    pipeline = PipelineManager.create(
        name="test", delta_root=tmp_path / "delta", staging_root=tmp_path / "staging",
    )
    output = pipeline.output

    pipeline.run(DataGenerator, name="gen", params={"count": 2, "seed": 42})
    pipeline.run(TransformAndScore, name="ts",
        inputs={"data": output("gen", "datasets")})

    summary = pipeline.finalize()
    assert summary.steps_completed == 2

Integration test (expanded)

def test_composite_expanded(tmp_path):
    pipeline = PipelineManager.create(
        name="test", delta_root=tmp_path / "delta", staging_root=tmp_path / "staging",
    )
    output = pipeline.output

    pipeline.run(DataGenerator, name="gen", params={"count": 2, "seed": 42})
    expanded = pipeline.expand(TransformAndScore, name="ts",
        inputs={"data": output("gen", "datasets")})

    summary = pipeline.finalize()
    # Expanded creates one step per internal operation
    assert summary.steps_completed >= 3

---

Style Rules

Follow these conventions (mirrors write-operation):

• Module docstring: One line, describes what the composite does
• Class docstring: Summary line + optional extended description
• Section comments: Use # ---------- Section ---------- with exactly 10 dashes on each side
• Section order: Metadata, Inputs, Outputs, Parameters, Compose
• compose() docstring: One-line imperative summary
• name value: snake_case matching the class name's snake_case form
• Imports: Group stdlib, then artisan composites, then artisan operations. Use from __future__ import annotations
• No bare constants: Put configurable values in a Params class or as params in ctx.run() calls