workflow-engine

name	workflow-engine
version	1.0.0
description	Use when the user asks to run a reusable YAML workflow, says "run the [workflow-name] workflow", "analisa ini dari 3 sudut", "debug this flaky test", "self-improve", "incident analysis", or "stress test business plan". Loads workflow definitions from ~/.openclaw/workflows and executes them safely with subagents.

Workflow Engine

Run reusable OpenClaw workflow definitions stored in ~/.openclaw/workflows/*.yaml.

Triggers

Use this skill when the user says any of:

• run the [workflow-name] workflow
• analisa ini dari 3 sudut
• debug this flaky test
• self-improve
• incident analysis
• stress test business plan

Files

• Workflow definitions: ~/.openclaw/workflows/*.yaml
• Safe helper CLI: /root/.openclaw/workspace-radit/scripts/workflow_runner.py

Safety rule

Do not rely on the runner to call tools or spawn agents. The runner only lists, validates, renders, and prints execution plans/prompts. Script steps are unsafe by default and only run from the CLI with explicit --allow-scripts; do not use that unless the user explicitly asked for local script execution and the command is safe.

Workflow execution procedure

1. Identify the workflow name.
   • If no exact workflow exists, list available workflows with: python3 scripts/workflow_runner.py list
   • If the workflow is missing, offer to create a new YAML workflow.
2. Read the workflow YAML from ~/.openclaw/workflows/<name>.yaml.
3. Validate it: python3 scripts/workflow_runner.py validate <name>
4. Collect required inputs from the YAML inputs mapping. If the user already supplied enough context, proceed without asking.
5. Render or run the plan for visibility:
   • Render prompts: python3 scripts/workflow_runner.py render <name> --input key=value
   • Safe run plan: python3 scripts/workflow_runner.py run <name> --input key=value
6. Execute steps manually as OpenClaw subagents:
   • Steps with no depends_on can run in parallel.
   • Respect depends_on and only start dependent steps after required results are back.
   • Use sessions_spawn for subagent steps when available.
   • Use max 3 concurrent subagents. If a batch has more than 3 independent steps, run it in waves.
   • Put the rendered step prompt in the subagent task.
7. Collect subagent outputs and synthesize the final answer for the user using the workflow output field and any final/synthesis step.
8. If the workflow contains script steps, present the command and ask before execution. Prefer a plan over execution.

Dependency handling

• Missing depends_on means independent and eligible for the current parallel batch.
• depends_on: [a, b] means wait for results from a and b, then expand any ${a} / ${b} placeholders with their outputs.
• For the included business-stress-test workflow, run investor, customer, and competitor in parallel, then run synthesize after those three results are available.

Included workflows

• flaky-test: debug flaky tests and produce root-cause/fix/verification plan.
• self-improve: review an agent/process/skill/artifact and produce prioritized improvements.
• incident-analysis: timeline, impact, root-cause hypotheses, mitigation, and prevention.
• business-stress-test: investor/customer/competitor parallel analysis followed by synthesis.