Documentation Index
Fetch the complete documentation index at: https://docs.igent.ai/llms.txt
Use this file to discover all available pages before exploring further.
Common issues and solutions for working with Maestro.
Need Help?
If you can’t find a solution to your issue in this troubleshooting guide, reach out to our support team:
Email: support@igent.ai
We’re here to help you get the most out of Maestro.
Session and Capacity Issues
”Session capacity is full”
Symptoms:
- Warning about token usage
- Maestro mentions capacity constraints
- Degraded performance
Root causes:
- Long conversation history
- Many files in view
- Multiple large files
- Accumulated iteration history
Solutions (in order of least to most aggressive):
-
Refresh files to latest only:
/refresh
Immediate relief if many file iterations
-
Hide experimental iterations:
/hidefiles
Select old/experimental iterations to hide
-
Compact memories:
/compact
Summarizes old dialog, preserves key decisions
-
Forget unnecessary turns:
/forget
Remove turns no longer needed
-
Start new session:
If session too large, consider fresh start:
- /synopsis to document current session
- Create new session
- Upload synopsis
- Continue with clean capacity
Prevention:
- Manage capacity proactively
- Use
/refresh regularly
- Don’t accumulate unnecessary iterations
- Create focused sessions
”Session won’t load” or “Loading timeout”
Causes:
- Very large session state
- Network connectivity issues
- Server-side restoration issues
Solutions:
- Refresh browser page
- Clear browser cache
- Check internet connection
- Try different browser
- Contact support if persists
For very large sessions:
- May take 30-60 seconds to load
- Be patient during restoration
- Consider session size management
Session state seems wrong
Symptoms:
- Files you created are missing
- Changes disappeared
- Conversation history incomplete
Causes:
- Browser cache issues
- Multiple tabs/sessions confusion
- Checkpoint restoration issue
Solutions:
-
Refresh browser
-
Check you’re in correct session
-
Use file restore if needed:
"Restore auth.py to iteration 10"
Maestro uses Restore Files tool
-
Check iteration history:
"Show me all iterations of auth.py"
Verify expected iterations exist
File and Code Issues
”File not found” when Maestro tries to view
Causes:
- Typo in filepath
- File in wrong directory
- File not created yet
- Case sensitivity issue
Solutions:
1. List files to find correct path:
"List all Python files in the project"
2. Search for file by content:
"Search all files for 'class UserService'"
3. Check exact spelling and case:
Filepaths are case-sensitive
src/Main.py ≠ src/main.py
Proposals not applied
Symptoms:
- Maestro created code proposals
- Files don’t reflect changes
- Sandbox doesn’t have new code
Cause: Apply Changes not invoked
Solution:
Maestro MUST use Apply Changes tool after all proposals
If you see proposals but no changes:
"Apply the pending changes to the codebase"
Maestro will invoke Apply Changes tool
Code changes disappeared
Likely cause: Working on wrong iteration
Check:
"What iteration of auth.py am I viewing?"
"Show me all iterations of auth.py with timestamps"
Solution:
View correct iteration:
"View auth.py iteration 15"
Or restore if needed:
"Restore auth.py to iteration 15"
Files not syncing to sandbox
Check:
- Is file under synchronized paths?
- Is file type synchronized (text-based)?
- Is file .gitignored?
Solutions:
Ask Maestro to:
"Check if myfile.py is in the sandbox at /home/sandbox/src/myfile.py"
"Move myfile.py to /home/sandbox/src/ in the sandbox"
"Check if myfile.py is in .gitignore"
"Show me the contents of myfile.py in the sandbox"
Sandbox Issues
Sandbox not responding
Symptoms:
- Commands timeout
- No terminal output
- “Sandbox unavailable” message
Solutions:
-
Wait for auto-recovery:
- Sandbox health monitoring active
- Automatic remediation attempts
- Usually recovers in 30-60 seconds
-
Force reset:
/reset-sandbox
Confirms your intent
Terminates and recreates sandbox
Files preserved (they're in session)
-
Check resource consumption:
Ask Maestro to:
"Check CPU and memory usage in the sandbox with htop"
"Find any runaway Python processes"
"Kill process [PID] if it's consuming too many resources"
Command hangs or times out
Causes:
- Command genuinely taking long time
- Process waiting for input
- Infinite loop or deadlock
Solutions:
-
For expected long commands:
Be patient
Check terminal stream for progress
Use Check Terminal Window to monitor
-
For hung commands:
Use Terminal Operator to send Ctrl+C
Or close terminal window
Or /reset-sandbox if unresponsive
-
Prevention:
For potentially long commands, ask Maestro to:
"Run this command with a 5-minute timeout: [command]"
"If the command takes longer than 5 minutes, cancel it"
Port 8080 already in use
Cause: Previous server not stopped
Solution:
Ask Maestro to:
"Find what process is using port 8080"
"Kill the process using port 8080"
Or:
"Start the server in a different terminal window"
"Use port 3000 instead of 8080"
Package installation fails
Common issues:
Permission denied:
Error: Permission denied
Solution: Ask Maestro to "Install [package] with sudo using apt-get"
Example: "Install python3-pip with sudo"
Package not found:
Solutions: Ask Maestro to:
"Update apt package database, then install [package]"
"Check if [package] is the correct package name for Ubuntu"
"Search for [package] in the Ubuntu repositories"
Note: Sandbox runs Ubuntu which uses apt-get package manager.
Common causes and solutions:
Invalid tool input:
Cause: Validation error in JSON
Solution: Maestro should see error and retry with corrected input
If stuck: Point out the validation error
Tool not available:
Cause: Tool disabled for session
Check: /tools to see available tools
Solution: Enable tool if needed
Resource constraints:
Cause: Sandbox out of memory/CPU
Solution:
- Free resources (kill processes)
- Use lighter-weight approach
- Split into smaller operations
Merge safety warnings
What are they:
Maestro proposes changes but hasn't applied them yet
Tries to use another tool
System detects potential for lost proposals
Solutions presented:
-
Apply Changes first (recommended):
- Commits pending proposals
- Then runs intended tool
-
Execute anyway:
- Proceeds with tool
- Proposals still pending
- Acceptable for read-only tools
-
Cancel and provide feedback:
- Cancels tool execution
- You explain what to do instead
Best practice: Apply changes before using other tools to avoid warnings.
Maestro will tell you:
Tool invocation has validation error:
[Specific error message]
Cross reference with tool schema and correct the input.
What to do:
- Read the error carefully
- Check tool schema for correct parameters
- Verify JSON structure
- Maestro should auto-correct on next attempt
If Maestro keeps failing:
"Stop trying that approach. Let's use a different tool or manual method."
Web and Network Issues
”Failed to retrieve URL”
Causes:
- URL requires authentication
- URL is localhost (not accessible from web tools)
- Rate limiting
- Server down
Solutions:
- For localhost: Use sandbox tools instead
- For auth: Provide secrets if possible
- For rate limits: Wait and retry
- For down servers: Verify URL is correct
Causes:
- URL not publicly accessible (localhost, private network)
- Site blocks automated browsers
- JavaScript rendering errors
Solutions:
- For localhost: Deploy to preview URL first, then screenshot
- For blocked sites: Some sites detect automation; may not work
- For render errors: Check browser console logs
Browser operator stuck
Symptoms:
- Browser operator not completing
- Timeout errors
- Wrong page state
Solutions:
- Simplify the goal
- Break into smaller steps
- Use explicit navigation instructions
- Try screenshot + manual analysis instead
Source Control Issues
Clone failed
Repository too large:
Error: Repository exceeds size limit
Current limit: ~10GB
Solution: Contact support or clone subset
Authentication failed:
Error: Unable to authenticate
Solution:
- Verify GitHub account linked
- Check repository exists
- Ensure you have access
Branch not found:
Error: Branch 'develop' does not exist
Solution:
- Verify branch name (case-sensitive)
- Try default branch
- Check branch exists on remote
PR creation failed
No changes detected:
Cause: No files modified or selected
Solution:
- Verify files were changed
- Ensure Apply Changes ran
- Select changed files in PR UI
Branch conflict:
Error: Branch 'feature/auth' already exists
Solutions:
- Use different branch name
- Update existing PR instead
- Delete remote branch first (if safe)
Merge conflict:
Error: Cannot merge due to conflicts
Current limitation: No automated resolution
Solution:
- Pull latest changes (overwrite clone)
- Manually resolve via Maestro guidance
- Or push directly and resolve on GitHub
Test and Validation Issues
Tests won’t run
Common causes:
Dependencies missing:
Error: Module not found
Solution: Ask Maestro to:
"Install Python dependencies from requirements.txt"
"Install npm dependencies from package.json"
Wrong Python environment:
Solution: Sandbox uses system Python with default venv
No need to create separate environment
If environment issues: "Verify Python environment setup in the sandbox"
Test framework not installed:
Solution: Ask Maestro to:
"Install pytest for running Python tests"
"Install jest as a dev dependency for JavaScript tests"
Tests failing unexpectedly
Investigation process:
1. Read full error output carefully
2. Understand what test expects
3. Check if test or implementation is wrong
4. Fix the correct component
5. Verify all tests pass
Don’t let Maestro:
- Skip failing tests
- Comment out tests
- Change tests to match wrong implementation
Do require Maestro to:
- Understand why tests fail
- Fix root cause
- Ensure all tests pass
After optimization attempt:
Investigate:
"The optimization made performance worse. Let's understand why:
1. Profile both implementations
2. Compare execution paths
3. Identify what caused slowdown
4. Revert change
5. Try different optimization approach"
Verification:
Always compare to baseline
Use identical test conditions
Run multiple iterations
Account for variance
Memory and Context Issues
”Can’t remember something we discussed”
Causes:
- Information was forgotten/compacted
- Happened many turns ago
- Context limit reached
Solutions:
Provide the information again:
"As I mentioned earlier, we're using PostgreSQL not MySQL"
Or reference files:
"Check the DATABASE_URL in config.py for database type"
Or use synopsis:
"/synopsis to review what we've done"
Maestro misunderstands context
When Maestro acts on wrong assumption:
Immediate correction:
"Stop. That's not correct. [Explain the actual situation]. Please re-approach based on this correct understanding."
Maestro will course-correct immediately
Prevention:
- Be explicit about constraints upfront
- Provide context in initial request
- Correct misunderstandings early
Download and File Export Issues
Downloaded files have wrong structure
Cause: Extraction at wrong location
Solution:
For /download-changed:
Extract zip at your PROJECT ROOT (not inside a subdirectory)
Example workflow:
1. Download the zip file
2. Navigate to your project root: cd ~/my-project
3. Extract there: unzip maestro_changes_*.zip
Files will slot into their correct locations relative to project root
Zip file seems incomplete
Check:
- Did you use correct download type?
- Are files .gitignored?
- Were files actually changed?
Verify:
Before download:
"List all changed files"
Verify expected files are present
Then download to confirm all included
Can’t download large session
Current behavior:
- All downloads should work
- Large sessions may take time
If timeout occurs:
- Use /download-changed (smaller subset)
- Or /download-recent (just latest files)
- Or specify specific files
Secrets and Integration Issues
Quick reference:
- Secret not working → Check registration and activation
- OAuth expired → Refresh in Secret Manager
- Permission denied → Check scopes/permissions
- Accidentally committed → Rotate immediately
Getting Help
Self-Service
-
Ask Maestro:
"I'm having trouble with [X]. How do I fix this?"
Maestro has extensive knowledge of its own system
-
Search documentation:
Use documentation search
Check relevant section (Sandbox, Source Control, etc.)
-
Review examples:
Working examples often clarify confusion
Check Quick Start and Session Strategies
Community and Support
Effective Bug Reports
Include:
- What you were trying to do
- What you expected to happen
- What actually happened
- Steps to reproduce
- Session ID (if applicable)
- Screenshots or error messages
- Your plan (if relevant)
Example:
Issue: Sandbox won't start
Expected: Sandbox starts when I run first command
Actual: Timeout error after 60 seconds
Steps:
1. Start new session
2. Request: "Install Python package xyz"
3. Maestro tries Execute Command
4. Times out
Session ID: abc-123-def-456
Error: [paste error message]
Plan: Basic (if relevant)
Common Misconceptions
”I broke Maestro”
Reality: Almost everything is recoverable:
- Files have iteration history (restore any version)
- Sessions checkpoint automatically
- Sandbox can be reset
- Secrets are isolated
Recovery tools:
- Restore Files
/reset-sandbox
- Reload session
- Create new session
”Maestro forgot everything”
Reality: Information persists in different forms:
- Dialog in memories
- Decisions in file content
- Clone records (never forgotten)
- Iteration history
Access forgotten context:
View old iterations for implementation details
Clone records show which repos/branches
File diffs show what changed and when
“This will take forever to fix”
Reality: Maestro is designed for iteration:
- Fast failure and correction cycles
- Can try multiple approaches quickly
- Learns from failures
- Often finds solutions through systematic exploration
Approach:
"This isn't working. Let's step back:
1. What have we tried?
2. What failed and why?
3. What haven't we tried?
4. New approach recommendation"
Maestro will synthesize and propose better path
Debugging Maestro Itself
Maestro seems confused
Symptoms:
- Contradictory statements
- Repeating same mistakes
- Not incorporating feedback
Solutions:
-
Clear correction:
"Stop. You're misunderstanding X. Here's the correct information: [clear explanation]"
-
Reset context:
"/forget last 3 turns"
Removes confusing interaction
Start fresh with clarity
-
Provide explicit constraints:
"For this implementation:
- Use Python 3.11 (not 3.9)
- PostgreSQL (not MySQL)
- Pytest (not unittest)
Confirm you understand before proceeding."
Maestro making assumptions
When Maestro assumes wrong things:
** Don't let pass:
Maestro: "Using MySQL..."
(But you want PostgreSQL)
** Correct immediately:
You: "Stop. We're using PostgreSQL, not MySQL. Update all database code accordingly."
Prevention:
- State constraints explicitly
- Provide all relevant context upfront
- Correct assumptions immediately
Maestro asks questions instead of researching
Sometimes appropriate:
- User-specific preferences
- Business requirements
- Architectural choices
Sometimes inappropriate:
- Technical information (use Perplexity Search)
- API documentation (use web tools)
- Best practices (use research tools)
When inappropriate:
You: "Use Perplexity Search to research the answer to your question, then proceed"
Maestro should use tools before asking
Maestro seems slow
Expected behavior:
- Complex reasoning takes time
- Tool execution has inherent latency
- Validation is thorough (not instant)
Genuine slowness:
Check:
- Is session capacity very high? (Affects token processing)
- Is sandbox under heavy load? Ask Maestro to "Check sandbox resource usage"
- Is network slow? (Affects web tools)
Solutions:
- Manage capacity (/refresh, /compact)
- Free sandbox resources: "Kill any unnecessary processes in the sandbox"
- Use faster tools when appropriate
Long response times
For research-heavy requests:
Expected:
- Perplexity Search (research mode): ~30 seconds
- Complex Reasoning: 1-3 minutes
- Browser Operator: Variable based on task
If longer than expected:
- Check if Maestro is stuck
- Cancel and simplify request
- Break into smaller parts
When Nothing Works
Nuclear Options
Last resort solutions (use sparingly):
-
Reload session:
Refresh browser completely
Clear browser cache if needed
Session state should restore
-
New session:
If session truly corrupted:
- /synopsis to document current state
- /download-all to preserve files
- Start new session
- Upload files and synopsis
- Resume work
-
Contact support:
For persistent issues:
- Document problem thoroughly
- Include session ID
- Describe steps to reproduce
- Contact support@igent.ai
Prevention Better Than Cure
Proactive Practices
Regular maintenance:
Every 20-30 turns:
- /refresh to clean file view
- Review capacity bar
- Consider /compact if high
Every major milestone:
- /synopsis to document
- /download-changed for backup
Clear communication:
Upfront:
- State all constraints
- Provide complete context
- Define success criteria
- Set quality standards
During:
- Correct misunderstandings immediately
- Provide feedback frequently
- Challenge assumptions early
After:
- Validate thoroughly
- Document learnings
- Package deliverables
Defensive validation:
Never skip:
- Running tests
- Checking benchmarks
- Verifying edge cases
- Reviewing all changes
Cost of validation: Minutes
Cost of undetected issues: Hours or days
Next Steps
Troubleshooting mastered: