ARK SDK and Type Development

Type Generation Pipeline

Types flow through the ARK stack:

CRD YAML (ark/config/crd/) ↓ make ark-sdk-build (automatic) ark-sdk Python types + overlay utilities ↓ ark-api Pydantic models (manual updates if exposing new fields) ↓ make ark-api-build → generates openapi.json ark-dashboard TypeScript types ↓ npm run generate:api → generates types.ts

ark-sdk: Automatic Type Generation

CRD types are automatically generated from Kubernetes CRD YAML files.

make ark-sdk-build

This runs:

• crd_to_openapi.py

• converts CRDs to OpenAPI schema

• openapi-generator-cli

• generates Python types from schema

• Copies overlay files on top of generated code

• Builds wheel package

Adding Custom SDK Functionality (Overlays)

The lib/ark-sdk/gen_sdk/overlay/python/ark_sdk/ directory contains custom Python modules that are copied on top of generated code. This is how utilities like client.py , k8s.py , and executor.py are added.

To add custom SDK functionality:

• Add Python files to lib/ark-sdk/gen_sdk/overlay/python/ark_sdk/

• Run make ark-sdk-build

ark-api: Manual Model Updates

The ark-api has manually written Pydantic models in services/ark-api/ark-api/src/ark_api/models/ .

When to update ark-api models:

• When exposing new CRD fields via the HTTP API

• When adding new API response types

make ark-api-build

This generates services/ark-api/openapi.json from FastAPI routes and Pydantic models.

Pydantic Model Naming (CRITICAL)

Pydantic model class names MUST be globally unique across all model files. When multiple files define classes with the same name (e.g., ConfigMapKeyRef , Header , ValueFrom ), FastAPI generates non-deterministic OpenAPI schema names like ark_api__models__agents__Header-Input . These names depend on import order and cause CI failures when types.ts differs between environments.

Solution: Prefix class names with their domain context:

File Naming Pattern Examples

agents.py

Agent*

AgentHeader , AgentValueFrom , AgentParameter

evaluators.py

Evaluator*

EvaluatorParameter , EvaluatorValueSource

mcp_servers.py

MCPServer*

MCPServerHeader , MCPServerValueSource

queries.py

Query*

QueryParameter , QueryLabelSelector

models.py

Model*

ModelValueSource

Safety net: generate_openapi.py fails if any schema name contains models , catching collisions at build time.

See: https://github.com/mckinsey/agents-at-scale-ark/issues/656

ark-dashboard: TypeScript Type Generation

cd services/ark-dashboard/ark-dashboard cp ../../ark-api/openapi.json ../out/ npm run generate:api npm run build # verify types compile

Debugging Type Errors

When you see:

Property 'SomeSchema' does not exist on type

• Regenerate types.ts from latest openapi.json

• Find the correct schema name: grep "SomeSchema" services/ark-dashboard/ark-dashboard/lib/api/generated/types.ts

• Update service files to use correct schema names

If you see schema names with models pattern (e.g., ark_api__models__agents__Header-Input ):

• This indicates a Pydantic model naming collision

• Fix by renaming the Python class to be unique (see naming convention above)

• Do NOT work around by using the models name - fix the root cause

Key Files

File Purpose

ark/config/crd/bases/*.yaml

Source of truth - Kubernetes CRDs

lib/ark-sdk/gen_sdk/overlay/

Custom SDK utilities (copied on top of generated)

services/ark-api/ark-api/src/ark_api/models/

Manually written Pydantic models

services/ark-api/ark-api/generate_openapi.py

Generates openapi.json with collision safety net

services/ark-api/openapi.json

Generated OpenAPI spec from FastAPI

services/ark-dashboard/out/openapi.json

Copy used for dashboard type generation

services/ark-dashboard/ark-dashboard/lib/api/generated/types.ts

Generated TypeScript types