workflow

name	workflow
description	Create and manage multi-step automation workflows combining any tools (ADB, shell, web, file) with variable interpolation. Supports standalone CLI execution, cron scheduling, and scripting.
metadata	{"pepebot":{"emoji":"🔄","requires":{},"platform":"all"}}

Workflow Creation Skill

JSON Structure

{
  "name": "Workflow Name",
  "description": "What it does",
  "variables": {
    "device": "",
    "param": "default_value"
  },
  "steps": [
    {
      "name": "step_id",
      "tool": "tool_name",
      "args": {
        "param": "{{variable}}"
      }
    },
    {
      "name": "analysis",
      "goal": "Instruction for LLM to interpret"
    },
    {
      "name": "use_skill",
      "skill": "skill_name",
      "goal": "Goal that uses the loaded skill content"
    },
    {
      "name": "delegate",
      "agent": "agent_name",
      "goal": "Goal to delegate to another agent"
    }
  ]
}

Rules

1. Every tool step MUST have "args": {} - even if empty. No exceptions.
2. Every step needs "name" - use snake_case identifiers.
3. 4 step types: tool+args, goal only, skill+goal, agent+goal.
4. tool cannot combine with skill or agent; skill and agent are mutually exclusive.
5. skill and agent require goal.
6. Use {{variable}} for interpolation in args and goals.
7. Step outputs are auto-stored as {{step_name_output}}.

IMPORTANT (workflow context only): When building workflow JSON and user says "use skill X" or "with skill X", use a skill step ("skill": "X", "goal": "..."). Do NOT manually replicate the skill's commands via tool/exec steps. The skill step automatically loads the skill content and combines it with the goal.

Do not interpret normal chat requests like "panggil agent", "switch agent", or "use agent" as an instruction to create/save a workflow unless the user explicitly asks for a workflow.

Available Tools

Tool	Required Args	Description
adb_devices	(none)	List connected devices
adb_shell	command, device(opt)	Run shell on device
adb_tap	x, y, device(opt)	Tap coordinates (numbers)
adb_input_text	text, device(opt)	Type text
adb_screenshot	filename, device(opt)	Capture screen
adb_ui_dump	device(opt)	Get UI XML hierarchy
adb_swipe	x1,y1,x2,y2, duration(opt), device(opt)	Swipe gesture
exec	command	Run shell command
read_file	path	Read file
write_file	path, content	Write file
web_search	query	Search web

Workflow Tools (Agent)

• workflow_save: Save workflow JSON → workflow_save(workflow_name, workflow_content)
• workflow_execute: Run workflow → workflow_execute(workflow_name, variables?)
• workflow_list: List all workflows → workflow_list()

Workflow CLI (Standalone)

Workflows can be executed directly from the terminal without the agent. This enables cron jobs, shell scripts, CI/CD, and headless automation.

# List workflows
pepebot workflow list

# Show workflow details (steps, variables, types)
pepebot workflow show <name>

# Run workflow from workspace
pepebot workflow run <name>

# Run with variable overrides (repeatable)
pepebot workflow run <name> --var device=emulator-5554 --var query=hello

# Run directly from any JSON file (bypass workspace)
pepebot workflow run -f /path/to/workflow.json
pepebot workflow run -f /path/to/workflow.json --var key=value

# Validate workflow structure
pepebot workflow validate <name>
pepebot workflow validate -f /path/to/workflow.json

# Delete a workflow
pepebot workflow delete <name>

Key: -f flag

The -f (or --file) flag runs a workflow from any file path, bypassing workspace lookup. This is useful for:

• Testing new workflows before saving to workspace
• Running workflows from external repos or shared directories
• CI/CD pipelines that store workflows alongside code
• One-off automation scripts

Standalone vs Agent Execution

Feature	CLI (pepebot workflow run)	Agent (workflow_execute)
Tool steps	Executed directly	Executed directly
Goal steps	Logged but not interpreted	LLM interprets and acts
Skill steps	Content loaded, stored as variable	Content loaded + LLM processes
Agent steps	Not available (no gateway)	Delegates to other agents
Variable overrides	--var key=value	variables parameter
Use case	Cron, scripts, CI/CD, headless	Chat, interactive, LLM-driven

Tip: For tool-only workflows, CLI execution is faster and doesn't need an LLM. Design workflows with pure tool steps for standalone automation, and add goal/agent steps only when LLM reasoning is needed.

Patterns

Device Health Check

{
  "name": "Device Health",
  "description": "Check battery, memory, storage",
  "variables": {"device": ""},
  "steps": [
    {"name": "battery", "tool": "adb_shell", "args": {"command": "dumpsys battery", "device": "{{device}}"}},
    {"name": "memory", "tool": "adb_shell", "args": {"command": "free -h", "device": "{{device}}"}},
    {"name": "storage", "tool": "adb_shell", "args": {"command": "df -h", "device": "{{device}}"}},
    {"name": "report", "goal": "Analyze battery, memory, storage data and generate health report"}
  ]
}

App Launch & Screenshot

{
  "name": "App Launcher",
  "description": "Launch app and capture screen",
  "variables": {"device": "", "package": "com.android.settings"},
  "steps": [
    {"name": "launch", "tool": "adb_shell", "args": {"command": "am start -n {{package}}/.Settings", "device": "{{device}}"}},
    {"name": "wait", "tool": "adb_shell", "args": {"command": "sleep 3", "device": "{{device}}"}},
    {"name": "capture", "tool": "adb_screenshot", "args": {"filename": "app_screen.png", "device": "{{device}}"}}
  ]
}

Nested Workflow (Workflow Calling Workflow)

{
  "name": "Full Test Suite",
  "description": "Run login, tests, and logout workflows in sequence",
  "variables": {"device": ""},
  "steps": [
    {"name": "login", "tool": "workflow_execute", "args": {"workflow_name": "login_flow"}},
    {"name": "test", "tool": "workflow_execute", "args": {"workflow_name": "feature_tests"}},
    {"name": "logout", "tool": "workflow_execute", "args": {"workflow_name": "logout_flow"}}
  ]
}

Agent Delegation

{
  "name": "Multi-Agent Task",
  "description": "Delegate research to another agent",
  "variables": {"topic": "AI trends"},
  "steps": [
    {"name": "research", "agent": "researcher", "goal": "Research {{topic}} and provide a summary"},
    {"name": "save", "tool": "write_file", "args": {"path": "research.txt", "content": "{{research_output}}"}}
  ]
}

Standalone Automation Recipes

Cron: Scheduled Device Health Check

# Run every hour — pure tool steps, no LLM needed
0 * * * * pepebot workflow run device_health --var device=emulator-5554

# Run daily report, save output to file
0 8 * * * pepebot workflow run daily_report --var device=abc123 > /var/log/pepebot/daily_$(date +\%F).log 2>&1

Shell Script: Multi-Device Batch

#!/bin/bash
# Run a workflow across multiple devices
for device in emulator-5554 emulator-5556 abc123; do
  echo "=== Running on $device ==="
  pepebot workflow run device_health --var device="$device"
done

Shell Script: Test Before Deploy

#!/bin/bash
# Validate and run before deploying
pepebot workflow validate smoke_test || exit 1
pepebot workflow run smoke_test --var device="$DEVICE" --var build="$BUILD_NUMBER"

CI/CD: Run Workflow From Repo

# Run a workflow checked into the repo (not from workspace)
pepebot workflow run -f ./ci/workflows/integration_test.json --var env=staging

Combine CLI + Script Output

#!/bin/bash
# Collect data with workflow, post-process with other tools
pepebot workflow run collect_metrics --var device=emulator-5554 > /tmp/metrics.txt
# Parse and send to monitoring
grep "battery" /tmp/metrics.txt | curl -X POST -d @- https://monitoring.example.com/api/metrics

Systemd Timer (Alternative to Cron)

# /etc/systemd/system/pepebot-health.timer
[Timer]
OnCalendar=*:0/30
# Every 30 minutes

# /etc/systemd/system/pepebot-health.service
[Service]
ExecStart=/usr/local/bin/pepebot workflow run device_health --var device=emulator-5554

Variable System

1. Default variables: Defined in "variables" field
2. Runtime overrides: Passed via workflow_execute variables parameter or --var flag in CLI
3. Step outputs: Auto-created as {{step_name_output}} or {{step_name}} (works for both tool and goal steps)
4. Goal raw text: {{step_name_goal}} contains the original goal text, NOT the LLM output — use {{step_name_output}} to get the LLM-processed result

Design Tips for Powerful Workflows

• Tool-only workflows run standalone without LLM — ideal for cron and scripts
• Goal/agent steps require the agent loop — use pepebot agent -m "run workflow X" for those
• Nest workflows using workflow_execute as a tool step to build modular test suites
• Use -f flag to iterate on workflow JSON files before saving to workspace
• Validate before deploying with pepebot workflow validate to catch errors early
• Log output by redirecting CLI output to files for audit trails
• Combine --var with environment variables in scripts: --var device="$DEVICE_SERIAL"

Notes

• Coordinate values (x, y, x1, y1, etc.) are auto-converted from strings to numbers
• Tool names and required parameters are validated on save
• adb_ui_dump may not work on all devices - use coordinates as fallback
• Workflow files are stored in ~/.pepebot/workspace/workflows/ as JSON