Overview Each Forge task can have multiple attempts - different executions with different AI agents, prompts, or approaches. This is Forge’s superpower: experiment until you find what works!
What Are Attempts? An attempt is a single execution of a task by an AI coding agent in an isolated Git worktree.
Task: "Add user authentication" ├── Attempt 1: Claude Sonnet → Too complex ├── Attempt 2: Gemini Flash → Missing edge cases ├── Attempt 3: Claude Haiku → Perfect! ✅ └── Result: You choose Attempt 3 to merge
Each attempt runs in complete isolation - no conflicts, no mess!
Creating Attempts First Attempt (via UI)
Create Task
Create a task in Forge with title and description
Click 'Start Task'
Click the task card → Start Task button
Choose Agent
Select your AI coding agent (Claude, Gemini, etc.)
Monitor Progress
Watch real-time logs as the agent works
Additional Attempts (via CLI) # Try the same task with a different agent forge task fork 1 --llm gemini # Try with a specialized agent profile forge task fork 1 --llm claude --agent "test-writer" # Try with different context forge task fork 1 --llm cursor --description "Focus on performance"
Via MCP You: "Try that authentication task again but with Gemini this time" Claude Code: I'll create a new attempt with Gemini... [Uses MCP to fork task] Attempt #2 started with Gemini
Attempt Lifecycle Status Explained
Status Meaning Actions Available Created Attempt queued, not started Start, Delete Running Agent is working Monitor, Cancel Completed Finished successfully Review, Merge, Fork Failed Execution failed Retry, View Logs Under Review Awaiting your approval Approve, Reject, Compare Merged Changes merged to main View, Rollback Rejected You decided not to use it Delete, Retry
Running Attempts Sequential Execution # Run one at a time forge task start 1 --llm claude # Wait for completion... forge task start 1 --llm gemini # Wait for completion...
Parallel Execution # Run multiple attempts simultaneously! forge task start 1 --llm claude & forge task start 1 --llm gemini & forge task start 1 --llm cursor & # All three run in isolated worktrees wait
Parallel execution is perfect for comparing approaches quickly!
Monitoring Attempts View Attempt Status # List all attempts for a task forge task attempts 1 # Output: # Attempt #1 (claude) [Completed] 5 minutes ago # Attempt #2 (gemini) [Running] Started 30s ago # Attempt #3 (cursor) [Failed] 2 minutes ago
Real-Time Logs # Follow logs as agent works forge task logs 1 --follow # View specific attempt forge task logs 1 --attempt 2 --follow
Web UI Monitoring The Forge UI shows live progress:
Current file being edited
Lines added/removed
Test results (if running)
Agent’s thinking process
Comparing Attempts See Comparing Results for detailed guide.
Quick comparison:
# Compare two attempts forge diff 1-claude 1-gemini # Compare all attempts forge task compare 1 # See just the stats forge task compare 1 --stats-only
Specialized Agent Profiles Use the same task with different “personas”:
Test Writer forge task fork 1 \ --llm claude \ --agent "test-writer" \ --description "Focus on comprehensive test coverage"
Security Expert forge task fork 1 \ --llm gemini \ --agent "security-expert" \ --description "Add input validation and sanitization"
forge task fork 1 \ --llm cursor \ --agent "performance-optimizer" \ --description "Optimize for speed, add caching"
Documentation Writer forge task fork 1 \ --llm claude \ --agent "docs-writer" \ --description "Add comprehensive documentation and examples"
Specialized agents are just different system prompts applied to the base LLM!
Each attempt stores:
{ "id" : "attempt-abc123" , "taskId" : "task-1" , "llm" : "claude" , "agent" : "default" , "status" : "completed" , "worktreePath" : ".forge/worktrees/task-1-claude" , "startedAt" : "2024-01-15T10:30:00Z" , "completedAt" : "2024-01-15T10:35:22Z" , "duration" : 322000 , "filesChanged" : 8 , "linesAdded" : 245 , "linesRemoved" : 12 , "cost" : { "inputTokens" : 15420 , "outputTokens" : 3890 , "totalCost" : 0.23 } }
View metadata:
forge task attempt-info 1 --attempt 2
Handling Failed Attempts View Failure Reason # Check logs forge task logs 1 --attempt 3 # Get error summary forge task attempt-info 1 --attempt 3 --errors-only
Common Failure Reasons
Issue : Agent generated invalid codeSolutions :
Retry with more specific instructions
Try different agent
Add example code in task description
Issue : Tests didn’t passSolutions :
Check test output in logs
Add “make tests pass” to description
Use --agent "test-aware" profile
Issue : Agent took too longSolutions :
Increase timeout: --timeout 600
Break task into smaller pieces
Use faster agent (Haiku, Flash)
Issue : LLM API failedSolutions :
Check API key is valid
Verify rate limits
Check network connectivity
Retry after a few minutes
Retry Failed Attempt # Retry with same configuration forge task retry 1 --attempt 3 # Retry with modifications forge task retry 1 --attempt 3 \ --llm gemini \ --description "Try simpler approach"
Cost Tracking View Attempt Costs # Cost for specific attempt forge task cost 1 --attempt 2 # Output: # Input tokens: 15,420 ($0.046) # Output tokens: 3,890 ($0.058) # Total cost: $0.104 # Total cost for all attempts forge task cost 1 --all # Output: # Attempt 1 (claude): $0.234 # Attempt 2 (gemini): $0.104 # Attempt 3 (cursor): $0.187 # Total: $0.525
Cost Optimization Strategies
Start Cheap Use fast, cheap models first:
Gemini Flash (free tier!)
Claude Haiku
GPT-3.5 Turbo
Save expensive models for complex tasks
Cancel Bad Attempts If agent is going wrong direction, cancel early: forge task cancel 1 --attempt 2
Saves tokens and money
Reuse Good Results Once you find an approach that works, save it as a template: forge template save 1 --attempt 2 \ --name "user-auth-pattern"
Track Spending Monitor total costs: forge cost summary --month january
Set budgets per project Attempt Best Practices
Start with One Agent
Don’t create 5 attempts immediately. Start with your preferred agent (Claude or Gemini)
Review Before More Attempts
If first attempt is close but not perfect, fork it with specific feedback rather than starting fresh
Use Different Agents for Different Strengths
Claude: Complex logic, architecture
Gemini: Fast iterations, simple tasks
Cursor: UI/UX focused work
Local models: Privacy-sensitive work
Clean Up Rejected Attempts
forge task cleanup 1 --remove-rejected
Keeps your workspace tidy Advanced: Attempt Hooks Run custom scripts on attempt events:
on_attempt_complete : - run : "npm test" - run : "npm run lint" - notify : "slack" on_attempt_fail : - run : "cat logs/error.log" - notify : "email"
Troubleshooting
Error : “Maximum attempts reached for task”Solution :
Default limit is 10 attempts per task
Clean up old attempts: forge task cleanup 1
Or increase limit: forge config set max-attempts 20
Error : “Worktree path already exists”Solution :# Clean orphaned worktrees forge worktree cleanup # Or remove specific worktree git worktree remove .forge/worktrees/task-1-old
Attempt stuck in 'Running'
Issue : Attempt shows running but nothing happeningSolution :# Cancel the attempt forge task cancel 1 --attempt 2 # Check for zombie processes ps aux | grep forge # Restart Forge forge restart
Next Steps 
