Workflow Orchestration

Workflow Orchestration Skill

Design and execute typed, multi-step workflows with approval gates.

Overview

The Lobster workflow engine enables deterministic pipelines that:

• Execute steps in sequence with state tracking
• Pause at approval gates for human verification
• Persist state across sessions
• Handle failures gracefully

Core Concepts

Workflow Definition

A YAML file defining the pipeline:

name: feature-pipeline
description: Complete feature implementation

steps:
  - name: plan
    command: /plan

  - name: implement
    agent: "@typescript-pro"

  - name: test
    command: /qa
    gate:
      type: manual
      message: "Tests pass. Continue?"

on_failure: notify

Step Types

Gate Types

Workflow Patterns

Pattern 1: Feature Development

name: feature-pipeline
steps:
  - name: plan
    command: /plan
    inputs:
      description: "{{ feature_description }}"

  - name: implement
    agent: "@typescript-pro"
    inputs:
      task: "Implement the plan"

  - name: test
    command: /qa
    gate:
      type: conditional
      condition: "test_results.failed == 0"

  - name: review
    agent: "@code-reviewer"
    gate:
      type: manual
      message: "Review complete. Ship?"

  - name: ship
    command: /ship

Pattern 2: Security Audit

name: security-audit
steps:
  - name: scan
    agent: "@security-auditor"
    outputs:
      - findings

  - name: analyze
    agent: "@zeno-analyzer"
    gate:
      type: conditional
      condition: "findings.critical == 0"
      fallback: fail

  - name: report
    shell: |
      echo "## Security Audit Report" > report.md
      # Generate report

  - name: notify
    command: /notify
    inputs:
      channel: "#security"

Pattern 3: Staged Deployment

name: deploy-pipeline
steps:
  - name: build
    shell: "npm run build"

  - name: test
    shell: "npm test"
    gate:
      type: conditional
      condition: "exit_code == 0"

  - name: deploy-staging
    command: /deploy-staging
    gate:
      type: timeout
      timeout_seconds: 300
      message: "Staging deployed. Auto-continues in 5 minutes."

  - name: smoke-test
    agent: "@test-automator"
    inputs:
      environment: staging

  - name: deploy-production
    shell: "npm run deploy:prod"
    gate:
      type: manual
      message: "Deploy to production?"
      approvers:
        - lead-developer
        - devops-team

Variables and Templating

Use {{ variable }} for dynamic values:

steps:
  - name: plan
    command: /plan
    inputs:
      description: "{{ feature_description }}"
      priority: "{{ priority }}"

Pass variables when starting:

/workflow feature-pipeline --var feature_description="Add OAuth" --var priority=high

State Management

Workflow state is persisted in .claude/artifacts/workflow-state/:

workflow-state/
├── feature-pipeline-abc123.json       # Workflow state
├── feature-pipeline-abc123.approval.json  # Pending approval
└── feature-pipeline-abc123.lock       # Concurrency lock

State Fields

Error Handling

Retry Configuration

steps:
  - name: flaky-test
    shell: "npm run e2e"
    retry_count: 3
    timeout_seconds: 300

Continue on Failure

steps:
  - name: optional-lint
    shell: "npm run lint"
    continue_on_failure: true

Failure Actions

on_failure: notify  # Send notification
on_failure: rollback  # (not yet implemented)
on_failure: continue  # Continue anyway

Best Practices

1. Gate Critical Steps: Always add manual gates before deploy/ship
2. Use Conditions: Auto-approve non-critical gates based on metrics
3. Set Timeouts: Prevent workflows from hanging indefinitely
4. Persist State: Let workflows resume across sessions
5. Notify on Gate: Alert team when approval is needed

Integration Points

Activation Triggers

This skill activates when prompts contain:

• "workflow", "pipeline", "orchestrat"
• "approval gate", "multi-step"
• "lobster", "typed workflow"
• "pause for approval", "wait for review"