Skip to main content

Add a workflow

A new customer process should be a workflow, not a fork of the kit. There are two contracts — pick the mode the project was initialized with.

Kit mode (default)​

  1. Write .pipeline/skills/{name}/SKILL.md (plus assets/ templates if needed).
  2. Write .pipeline/workflows/{name}.json with context.parent.files and context.steps.{agent}.files (relative paths from the project root, .pipeline/… prefix).
  3. Add workflows.{name} in .pipeline/config.json (source, chain or classes, skips). An empty chain means parent-only (see ask).
  4. Optional: add .pipeline/agents/{role}.md and list it on that step’s allowlist.
  5. If the receptionist should auto-pick it, add one row to .pipeline/skills/orchestration/SKILL.md (O1 / O3).
  6. Activate and check the printed list:
python3 .pipeline/loader/load_workflow.py --workflow {name} --step parent --slug try-{name}

Confirm allowed_reads is the smallest set that step needs.

Follow DOCUMENT-STANDARD so the new markdown stays portable.

Kit-repo copy of this contract: extensions/kit/.

Orchestrator mode​

--mode orchestrator does not use the markdown loader. Add a Python graph:

pipeline-kit workflows --scaffold my-review

That writes pipeline_extensions/my_review.py and a brief. Copy-ready examples (security review, CI audit, dependency audit, accessibility review) live in the kit clone at extensions/orchestrator/.

cp -R extensions/orchestrator/pipeline_extensions /path/to/your-app/

Do not reuse a first-party name. Kit mode never loads pipeline_extensions/.

Full comparison: Kit vs orchestrator, Extensions.