Composite Tasks

​

How composite tasks work

​

1. AI Planning Phase

• AutoDev analyzes your request
• AI generates a task graph (DAG structure)
• Tasks are organized with dependencies
• Plan is saved for your review

​

2. Review and Start

• View the task graph visualization
• See all tasks and their dependencies
• Review the execution plan
• Click Start Execution to begin

​

3. Task Execution

• Initial tasks (no dependencies) become Ready
• You approve each ready task
• Approved tasks spawn unit tasks
• Tasks run in parallel when possible
• New tasks become ready as dependencies complete

​

4. Completion

• Composite task status: Completed
• All pull requests are created

​

Creating a composite task

1. Go to your AutoDev dashboard
2. Select your repository
3. Enter your complex project description
4. Check “Create as Composite Task”
5. (Optional) Configure defaults for all tasks:
   • AI Agent: Default agent for all nodes
   • Model Version: Default model for all nodes
   • Base Branch: Branch for all tasks
6. Click Create Task

​

Task graph structure

​

Nodes

• Pending (gray) - Waiting for dependencies
• Ready (blue) - Dependencies met, awaiting approval
• Approved (yellow) - Approved, about to start
• In Progress (orange) - Currently executing
• Completed (green) - Successfully finished
• Failed (red) - Encountered an error
• Blocked (dark gray) - Cannot proceed due to failed dependency

​

Edges

• Task A → Task B means B depends on A
• Task B starts only after Task A completes

​

Parallel execution

    Task A
   /      \
Task B    Task C  (B and C run in parallel)
   \      /
    Task D        (D waits for both B and C)

​

Approving tasks

1. Review the task’s auto-generated prompt
2. (Optional) Modify the prompt before approving
3. Click Approve to start execution
4. Task spawns a unit task and begins work

​

Auto-approval mode

1. Go to composite task details
2. Enable Auto-Approval
3. Ready tasks are automatically approved and executed
4. No manual intervention required

• You trust the AI-generated plan
• Tasks are well-defined and low-risk
• You want continuous progress without manual steps

• You want to review each task before execution
• Tasks require careful sequencing
• You want to modify prompts before execution

​

Managing composite tasks

​

Task details page

• Title and description
• Repository and branch
• Progress (X/Y tasks completed)
• Current status
• Creation time

• Interactive diagram of all tasks
• Color-coded status indicators
• Click nodes to see details
• View dependencies

• List of tasks ready to execute
• Option to modify prompts
• Approve or reject buttons

• Currently executing tasks
• Links to spawned unit tasks
• Real-time status

• Finished tasks
• Links to pull requests
• Merge status

• Failed tasks with error details
• Retry options
• Blocked tasks showing required dependencies

​

Editing task prompts

1. Task becomes ready
2. Review the AI-generated prompt
3. Click Edit Prompt
4. Make your changes
5. Approve to execute with modified prompt

1. Go to task graph visualization
2. Click on a task node (must be pending, ready, or blocked)
3. Edit the prompt and title
4. Save changes
5. Approve when ready

​

Retrying failed tasks

1. Review the error details
2. (Optional) Edit the task prompt
3. Click Retry
4. A new unit task is created with the same (or updated) configuration

​

Syncing composite tasks

1. Go to composite task details
2. Click Sync Composite Task
3. AutoDev checks all tasks and PRs:
   • Updates PR merge statuses from GitHub
   • Marks completed tasks when PRs are merged
   • Recalculates node readiness
   • Triggers auto-approval if enabled
   • Updates composite task progress

• Tasks updated count
• Nodes updated count
• PRs synced count
• Workflows triggered count

​

Task graph best practices

​

How AI creates the plan

1. Minimize rebasing - Structure for parallel execution
2. CI/Infrastructure first - Set up testing before features
3. Independent code paths - Tasks touch different files
4. Clear boundaries - Define interfaces early

• Phase 1: CI/Infrastructure (no dependencies)
• Phase 2: Database/Storage setup
• Phase 3: Core contracts (APIs, models)
• Phase 4: Parallel features
• Phase 5: Integration
• Phase 6: Polish (UI, performance, docs)

• Wide and shallow (more parallel, fewer sequential)
• Each task represents meaningful progress

​

Writing effective composite task prompts

Build a complete e-commerce checkout system including:
- Shopping cart with session persistence
- Payment processing with Stripe integration
- Order confirmation emails
- Admin order management dashboard
- Inventory updates after purchase

Migrate to microservices architecture using:
- Docker containers for each service
- API Gateway with Kong
- Event bus with RabbitMQ
- Shared PostgreSQL database
- CI/CD with GitHub Actions

Implement authentication following our architecture:
- Reference: @docs/auth-architecture.md
- Use our existing JWT library
- Follow security guidelines in @SECURITY.md

​

Deleting composite tasks

1. Cancels all incomplete unit tasks
2. Blocks all pending/ready task nodes
3. Cancels running GitHub Actions workflows
4. Removes the composite task

​

Example composite task

Build a user authentication system with email verification and OAuth

1. Setup CI/CD pipeline
   ↓
2. Create database schema and migrations
   ↓
   ├→ 3. Implement email authentication
   │     ├→ User registration
   │     ├→ Email verification
   │     └→ Password reset
   │
   ├→ 4. Implement OAuth integration
   │     ├→ Google OAuth
   │     ├→ GitHub OAuth
   │     └→ OAuth callback handling
   │
   └→ 5. Build frontend components
         ├→ Login/register forms
         ├→ OAuth buttons
         └→ Email verification UI
   ↓
6. Integration tests and documentation

​

Benefits of composite tasks

​

Coordinated execution

• Tasks build on each other
• Dependencies ensure correct order
• Parallel work speeds completion

​

AI-powered planning

• Optimal task breakdown
• Smart dependency management
• Efficient parallelization

​

Flexibility

• Modify prompts before execution
• Retry failed tasks
• Enable/disable auto-approval

​

Visibility

• See entire project plan upfront
• Track progress across all tasks
• Understand dependencies

​

Batch processing with consistent prompts

Review all AutoDev documentation pages one by one and fix awkward translations.
Check each supported language individually.
Create one PR per page. Only create PRs if translations need fixing.

• Plan separate tasks for each documentation page
• Apply the same review logic to each language version
• Create individual PRs only when improvements are needed
• Process pages in parallel for faster completion

• Parallelization: Multiple pages are processed simultaneously
• Focus: Each AI agent works on a single page, preventing the common issue of AI agents forgetting their tasks midway through complex work
• Isolation: Problems in one page don’t affect others

Make the translation of each AutoDev page smooth and natural.

Create one task per page. Each task should handle one page.
Each task should not create a PR if there are no translation issues.

Update all API endpoints to use the new authentication middleware.
Create one PR per endpoint group.
Ensure backward compatibility in each change.

​

When not to use composite tasks

• Task is straightforward and single-purpose
• No dependencies on other work
• Quick fix or small change
• One pull request is sufficient

​

Troubleshooting

​

Next steps

• Learn about Unit Tasks
• Understand PR Review Process
• Configure AI Agents