Skip to main content

Process Definitions: JavaScript Workflow Orchestration

Version: 1.1 Last Updated: 2026-01-26 Category: Feature Guide

In Plain English​

A process is a recipe that tells Babysitter what to do.

Just like a cooking recipe says "chop vegetables, then cook them, then serve" - a process says "research the codebase, then write tests, then implement, then verify."

You don't need to write processes to use Babysitter. The Process Library is the SDK-managed library under library/, and the current generated snapshot counts 2,239 JavaScript process files in the live repository tree.

When would you write a process?

  • You have a specific workflow your team follows
  • The pre-built processes don't match your needs
  • You want to customize how quality gates work

Tip for beginners: Start with pre-built processes. Once you're comfortable, come back here to learn how to create your own.

Overview​

Process definitions are JavaScript functions that orchestrate workflows in Babysitter. A process defines what tasks to execute, in what order, and how to handle results. The process function acts as the "brain" of your workflow, making decisions and coordinating execution while Babysitter handles state management, persistence, and resumability.

Why Use Process Definitions​

  • Deterministic Execution: Same inputs and journal produce the same execution path
  • Full JavaScript Power: Use loops, conditionals, and async/await for complex logic
  • Modular Design: Define reusable tasks and compose them into workflows
  • Event-Sourced: All state changes recorded for replay and debugging
  • Resumable: Workflows automatically resume from where they left off

Use Cases and Scenarios​

Scenario 1: Simple Build and Test Pipeline​

A basic process that builds and tests a project.

import { defineTask } from '@a5c-ai/babysitter-sdk';

export async function process(inputs, ctx) {
  // Build the project
  const buildResult = await ctx.task(buildTask, { target: inputs.target });

  // Run tests
  const testResult = await ctx.task(testTask, { suite: 'unit' });

  // Return results
  return {
    success: buildResult.success && testResult.success,
    build: buildResult,
    tests: testResult
  };
}

const buildTask = defineTask('build', (args, taskCtx) => ({
  kind: 'node',
  title: `Build ${args.target}`,
  node: {
    entry: 'scripts/build.js',
    args: ['--target', args.target]
  },
  io: {
    inputJsonPath: `tasks/${taskCtx.effectId}/input.json`,
    outputJsonPath: `tasks/${taskCtx.effectId}/result.json`
  }
}));

const testTask = defineTask('test', (args, taskCtx) => ({
  kind: 'node',
  title: `Run ${args.suite} tests`,
  node: {
    entry: 'scripts/test.js',
    args: ['--suite', args.suite]
  },
  io: {
    inputJsonPath: `tasks/${taskCtx.effectId}/input.json`,
    outputJsonPath: `tasks/${taskCtx.effectId}/result.json`
  }
}));

Scenario 2: CI Pipeline with Conditional Steps​

A more complex pipeline with parallel execution and conditional logic.

export async function process(inputs, ctx) {
  // Build first
  const buildResult = await ctx.task(buildTask, { target: 'app' });

  // Run lint and tests in parallel
  const [lintResult, testResult] = await ctx.parallel.all([
    () => ctx.task(lintTask, { files: buildResult.files }),
    () => ctx.task(testTask, { suite: 'smoke' })
  ]);

  // Conditional: request approval if issues found
  if (!lintResult.ok || !testResult.ok) {
    await ctx.breakpoint({
      question: 'Lint/tests failed. Continue anyway?',
      title: 'Quality Gate',
      context: {
        runId: ctx.runId,
        files: [{ path: 'artifacts/quality-report.md', format: 'markdown' }]
      }
    });
  }

  // Agent-based code review
  const review = await ctx.task(codeReviewAgentTask, { diffRef: buildResult.diffRef });

  return { success: true, review: review.summary };
}

Scenario 3: Multi-Phase Feature Development​

A comprehensive workflow with research, planning, implementation, and verification phases.

export async function process(inputs, ctx) {
  const { feature, targetQuality = 85, maxIterations = 5 } = inputs;

  // Phase 1: Research
  const research = await ctx.task(researchTask, { feature });

  // Phase 2: Planning
  const plan = await ctx.task(planningTask, { feature, research });

  // Breakpoint: Approve plan
  await ctx.breakpoint({
    question: `Review plan for "${feature}". Approve to proceed?`,
    title: 'Plan Review',
    context: { runId: ctx.runId, files: [{ path: 'artifacts/plan.md', format: 'markdown' }] }
  });

  // Phase 3: Implementation with quality convergence
  let iteration = 0;
  let quality = 0;

  while (iteration < maxIterations && quality < targetQuality) {
    iteration++;

    const impl = await ctx.task(implementTask, { feature, plan, iteration });
    const score = await ctx.task(scoreQualityTask, { impl });

    quality = score.overall;
    ctx.log(`Iteration ${iteration}: Quality ${quality}/${targetQuality}`);
  }

  // Phase 4: Final verification
  await ctx.breakpoint({
    question: `Quality: ${quality}. Approve for merge?`,
    title: 'Final Approval'
  });

  return { success: quality >= targetQuality, iterations: iteration, quality };
}

Step-by-Step Instructions​

Step 1: Create the Process File​

Create a JavaScript file with an exported process function.

Location: .a5c/runs/<runId>/code/main.js or a custom path

// main.js
export async function process(inputs, ctx) {
  // Your workflow logic here
  return { success: true };
}

Step 2: Define Tasks​

Tasks are the building blocks of your workflow. Define them using defineTask.

import { defineTask } from '@a5c-ai/babysitter-sdk';

export const myTask = defineTask('my-task', (args, taskCtx) => ({
  kind: 'node',
  title: `Execute ${args.action}`,
  node: {
    entry: 'scripts/my-script.js',
    args: ['--action', args.action]
  },
  io: {
    inputJsonPath: `tasks/${taskCtx.effectId}/input.json`,
    outputJsonPath: `tasks/${taskCtx.effectId}/result.json`
  }
}));

Step 3: Use the Process Context​

The ctx object provides intrinsics for orchestration.

export async function process(inputs, ctx) {
  // Execute a task
  const result = await ctx.task(myTask, { action: 'build' });

  // Request human approval (returns BreakpointResult)
  const approval = await ctx.breakpoint({ question: 'Approve?' });
  if (!approval.approved) {
    return { success: false, reason: approval.feedback };
  }

  // Sleep until a specific time
  await ctx.sleepUntil('2026-01-26T09:00:00.000Z');

  // Execute tasks in parallel
  const [a, b] = await ctx.parallel.all([
    () => ctx.task(taskA, {}),
    () => ctx.task(taskB, {})
  ]);

  // Log to the journal
  ctx.log('Workflow completed', { result });

  // Get current time (deterministic)
  const now = ctx.now();

  return { success: true };
}

Step 4: Create a Run​

Use the CLI to create a run with your process.

babysitter run:create \
  --process-id my-workflow \
  --entry ./code/main.js#process \
  --inputs ./inputs.json \
  --run-id "run-$(date -u +%Y%m%d-%H%M%S)" \
  --prompt "Run my custom workflow"

Step 5: Execute the Run​

Use the babysitter skill or CLI to drive execution.

# Via skill
claude "/babysitter:call run my-workflow"

# Via CLI iteration loop
while true; do
  RESULT=$(babysitter run:iterate "$RUN_ID" --json)
  STATUS=$(echo "$RESULT" | jq -r '.status')
  [ "$STATUS" = "completed" ] && break
  [ "$STATUS" = "failed" ] && exit 1
done

Configuration Options​

Process Function Signature​

async function process(inputs, ctx) {
  // inputs: Object - Initial inputs passed to the run
  // ctx: ProcessContext - Orchestration intrinsics
  // returns: any - Final output of the process
}

Task Definition Schema​

defineTask<TArgs, TResult>(id: string, impl: TaskImpl<TArgs>): DefinedTask<TArgs, TResult>
FieldTypeRequiredDescription
kindstringYesTask type: node, shell, agent, breakpoint
titlestringNoHuman-readable title
descriptionstringNoDetailed description
node.entrystringYes (for node)Path to Node.js script
node.argsstring[]NoCommand-line arguments
node.envobjectNoEnvironment variables
node.cwdstringNoWorking directory
node.timeoutnumberNoTimeout in milliseconds
shell.commandstringYes (for shell)Shell command to execute
agent.namestringYes (for agent)Agent name
agent.promptobjectYes (for agent)Agent prompt configuration
io.inputJsonPathstringNoPath for input JSON
io.outputJsonPathstringNoPath for output JSON
labelsstring[]NoLabels for categorization
execution.harnessstringNoPreferred harness CLI for task execution (internal-only)
execution.modelstringNoPreferred model for agent tasks
execution.permissionsstring[]NoPermission list for task execution (internal-only)

Process Context Intrinsics​

MethodDescription
ctx.task(taskDef, args, options?)Execute a task
ctx.breakpoint(payload)Request human approval, returns BreakpointResult. Supports routing via expert, tags, strategy, previousFeedback, and attempt fields.
ctx.sleepUntil(timestamp)Sleep until a specific time
ctx.parallel.all(thunks)Execute tasks in parallel
ctx.parallel.map(items, fn)Map items to parallel tasks
ctx.log(...args)Log to the journal
ctx.now()Get current time (deterministic)
ctx.runIdThe current run ID

Code Examples and Best Practices​

Example 1: Node Task Definition​

export const buildTask = defineTask('build', (args, taskCtx) => ({
  kind: 'node',
  title: `Build ${args.target}`,
  description: 'Compile and bundle the application',
  node: {
    entry: 'scripts/build.js',
    args: ['--target', args.target, '--effect-id', taskCtx.effectId],
    env: { NODE_ENV: 'production' },
    timeout: 300000  // 5 minutes
  },
  io: {
    inputJsonPath: `tasks/${taskCtx.effectId}/input.json`,
    outputJsonPath: `tasks/${taskCtx.effectId}/result.json`
  },
  labels: ['build', 'production']
}));

Example 2: Shell Task Definition​

export const lintTask = defineTask('lint', (args, taskCtx) => ({
  kind: 'shell',
  title: 'Run linter',
  description: 'Check code style and common issues',
  shell: {
    command: `npx eslint ${args.files.join(' ')} --format json --output-file tasks/${taskCtx.effectId}/result.json`
  },
  labels: ['quality', 'lint']
}));

Example 3: Agent Task Definition​

export const codeReviewAgentTask = defineTask('code-review', (args, taskCtx) => ({
  kind: 'agent',
  title: 'AI Code Review',
  description: 'LLM-based code review',
  agent: {
    name: 'code-reviewer',
    prompt: {
      role: 'senior software engineer',
      task: 'Review the code changes and provide feedback',
      context: {
        diffRef: args.diffRef
      },
      instructions: [
        'Check for bugs and security issues',
        'Review code quality and style',
        'Suggest improvements'
      ],
      outputFormat: 'JSON with summary, issues, and suggestions'
    },
    outputSchema: {
      type: 'object',
      required: ['summary', 'issues'],
      properties: {
        summary: { type: 'string' },
        issues: { type: 'array', items: { type: 'object' } },
        suggestions: { type: 'array', items: { type: 'string' } }
      }
    }
  },
  execution: {
    model: 'claude-sonnet-4-20250514',  // Preferred model for this task
  },
  io: {
    inputJsonPath: `tasks/${taskCtx.effectId}/input.json`,
    outputJsonPath: `tasks/${taskCtx.effectId}/result.json`
  },
  labels: ['agent', 'code-review']
}));

Execution Hints​

Task definitions support optional execution hints that influence how the orchestrator runs the task:

FieldTypeVisibilityDescription
execution.modelstringUniversalPreferred model for agent tasks (e.g., 'claude-sonnet-4-20250514')
execution.harnessstringInternal-onlyPreferred harness CLI for task execution
execution.permissionsstring[]Internal-onlyPermission list for the task execution environment
const heavyAnalysisTask = defineTask('heavy-analysis', (args, taskCtx) => ({
  kind: 'agent',
  title: 'Deep code analysis',
  agent: { /* ... */ },
  execution: {
    model: 'claude-opus-4-20250514',         // Use a more capable model
    harness: 'claude',                       // Internal: prefer Claude Code harness
    permissions: ['read', 'write', 'shell'], // Internal: required permissions
  },
}));

Example 4: Error Handling in Processes​

export async function process(inputs, ctx) {
  try {
    const result = await ctx.task(riskyTask, { data: inputs.data });
    return { success: true, result };
  } catch (error) {
    // Log the error
    ctx.log('Task failed', { error: error.message });

    // Request human intervention
    await ctx.breakpoint({
      question: `Task failed: ${error.message}. How to proceed?`,
      title: 'Error Recovery',
      context: {
        runId: ctx.runId,
        files: [{ path: 'artifacts/error-details.json', format: 'code', language: 'json' }]
      }
    });

    // Retry with different parameters
    const retryResult = await ctx.task(riskyTask, { data: inputs.data, retryMode: true });
    return { success: true, result: retryResult, retried: true };
  }
}

Example 5: Dynamic Task Selection​

export async function process(inputs, ctx) {
  const { taskType, config } = inputs;

  // Select task based on input
  let taskDef;
  switch (taskType) {
    case 'build':
      taskDef = buildTask;
      break;
    case 'test':
      taskDef = testTask;
      break;
    case 'deploy':
      taskDef = deployTask;
      break;
    default:
      throw new Error(`Unknown task type: ${taskType}`);
  }

  const result = await ctx.task(taskDef, config);
  return result;
}

Best Practices​

  1. Keep Processes Deterministic: Avoid random values or non-deterministic operations; use ctx.now() for timestamps
  2. Use Descriptive Task IDs: Task IDs should clearly indicate what the task does
  3. Handle Errors Gracefully: Implement error handling and recovery strategies
  4. Break Complex Workflows into Phases: Structure processes with clear phases for readability
  5. Document Process Purpose: Add comments explaining the workflow logic
  6. Use Labels for Categorization: Tag tasks with labels for filtering and organization
  7. Separate Task Definitions: Keep task definitions in separate files for reusability

Common Pitfalls and Troubleshooting​

Pitfall 1: Non-Deterministic Process Code​

Symptom: Process behaves differently on resume.

Cause: Using non-deterministic values.

Wrong:

const timestamp = Date.now();  // Non-deterministic
const id = Math.random().toString(36);  // Non-deterministic

Correct:

const timestamp = ctx.now().getTime();  // Deterministic
const id = `task-${ctx.runId}-${iteration}`;  // Deterministic

Pitfall 2: Missing Input/Output Paths​

Symptom: Task results not persisted or not found on resume.

Cause: Missing io configuration.

Solution:

defineTask('my-task', (args, taskCtx) => ({
  kind: 'node',
  node: { entry: 'scripts/task.js' },
  io: {
    inputJsonPath: `tasks/${taskCtx.effectId}/input.json`,   // Add this
    outputJsonPath: `tasks/${taskCtx.effectId}/result.json`  // Add this
  }
}));

Pitfall 3: Process Code Changed Between Iterations​

Symptom: Replay produces different results.

Cause: Process code modified after run started.

Solution:

  • Avoid modifying process code for in-progress runs
  • The SDK stores processRevision to detect changes
  • Create a new run if workflow logic needs to change

Pitfall 4: Task Function Not Exported​

Symptom: ReferenceError: task is not defined

Cause: Task function not exported or imported incorrectly.

Solution:

// In tasks.js
export const myTask = defineTask('my-task', ...);

// In main.js
import { myTask } from './tasks.js';

export async function process(inputs, ctx) {
  const result = await ctx.task(myTask, { /* args */ });
}

Pitfall 5: Incorrect Entry Point Syntax​

Symptom: Error: Cannot find module

Cause: Incorrect entry point format in run:create.

Correct syntax:

--entry ./code/main.js#process
       ^              ^
       path           export name (after #)

Pre-Built Workflows: Methodologies & Processes​

Don't start from scratch! Babysitter includes thousands of ready-to-use workflows:

Methodologies (38 directories in this repo snapshot) - Development Approaches​

High-level approaches you can apply to any project:

  • TDD Quality Convergence - Test-first with iterative quality improvement
  • GSD (Get Stuff Done) - Rapid 8-phase execution workflow
  • Spec-Kit - Specification-driven with governance
  • Domain-Driven Design - Strategic and tactical DDD patterns
  • And many more under library/methodologies/

Browse methodologies:

Domain Processes - Task-Specific Workflows​

Complete process definitions for specific domains:

DomainProcessesBrowse
Development and technical specializations837Browse →
Business domains490Browse →
Science & engineering domains551Browse →
Social sciences & humanities160Browse →

See the full catalog with descriptions in the Process Library.

Summary​

Process definitions are JavaScript functions that orchestrate workflows. Define reusable tasks, compose them with conditionals and loops, and let Babysitter handle state management. Keep processes deterministic for reliable replay and resumption. Use the full power of JavaScript while benefiting from event-sourced persistence.