fix: prevent merge conflicts and improve parallel agent isolation

[BasselBlal]

Description:
This PR improves the reliability of parallel agent execution by strictly isolating infrastructure files, preventing merge conflicts, and handling all file changes robustly. It simplifies the commit logic to ensure clean working directories and data integrity.

Key Changes:

1. src/execution/parallel.ts:

   • Fix: Refactored runtime file filtering to use the centralized DEFAULT_IGNORED list from sandbox.ts.
   • Reason: Replaces ad-hoc path checks with a robust, case-insensitive logic that correctly filters broad patterns like .ralphy (entire folder), matching sandbox creation logic.

2. src/execution/sandbox-git.ts:

   • Fix: Replaced catch-all git staging with a two-phase staging process using a hardened safeGitAdd helper:
     1. Batching: Processes files in chunks of 20 to prevent ENAMETOOLONG errors on Windows.
     2. Ignore Filtering: Pre-checks files with git check-ignore to prevent errors from git adding ignored files (e.g. .env, .turbo).
     3. Retries: Automatically retries on index.lock contention to handle high-concurrency parallel execution.
     4. Two-Phase Strategy: First stages explicitly tracked files, then catches untracked files (e.g. composer installs) via git status, providing redundancy against missed files.
   • Reason: The previous git add . approach was fragile, and standard git add crashed on large file sets or locked repos. The new robust approach guarantees data integrity even under heavy parallel load.

3. src/execution/sandbox.ts:

   • Fix: Updated DEFAULT_IGNORED to include the entire .ralphy directory and improved matchesPattern regex escaping.
   • Reason: Agents do not need access to the .ralphy folder (config is loaded by the runner), preventing infrastructure pollution. Added comprehensive unit tests (sandbox.test.ts).

4. src/execution/prompt.ts:

   • Fix: Removed instructions to update progress.txt.
   • Reason: Eliminates race conditions where agents attempted to modify the runner's tracking file.

Result:
Parallel execution now runs cleanly without merge conflicts or "dirty checkout" errors. Infrastructure files are fully isolated, and the commit strategy ensures zero data loss for agent-generated files (including package manager artifacts) while respecting isolation boundaries.

[vercel]

@BasselBlal is attempting to deploy a commit to the Goshen Labs Team on Vercel.

A member of the Team first needs to authorize it.

[dosubot]

Related Documentation

Checked 30 published document(s) in 1 knowledge base(s). No updates required.

^{How did I do? Any feedback?}  

[greptile-apps]

Greptile Overview

Greptile Summary

This PR successfully addresses parallel agent isolation by preventing infrastructure file conflicts through centralized ignore patterns and a robust two-phase git staging process.

Key Improvements

• Centralized Filtering: Introduced DEFAULT_IGNORED list with pattern matching functions (matchesPattern, isIgnored) that properly handle directory boundaries, wildcards, and regex escaping
• Infrastructure Isolation: .ralphy/ directory and other infrastructure files are now filtered from both sandbox creation and git commits, preventing merge conflicts
• Robust Git Staging: Two-phase approach with safeGitAdd helper that batches files (prevents ENAMETOOLONG), filters ignored files via git check-ignore, and retries on lock contention
• Race Condition Fix: Removed progress.txt update instruction from agent prompts to eliminate concurrent modification issues
• Comprehensive Tests: Added test suite validating pattern matching edge cases including boundary checks and regex escaping

Architecture

The filtering logic is applied consistently across three critical points:

1. Sandbox creation (sandbox.ts:189-199) - prevents infrastructure files from being copied into agent workspaces
2. File commit filtering (parallel.ts:470-506) - prevents infrastructure files from being staged for commit
3. Git staging (sandbox-git.ts:164-191) - secondary filter during two-phase staging catches any edge cases

The two-phase staging approach (Phase 1: explicit filtered files, Phase 2: untracked files from git status) ensures no agent-generated files are lost while maintaining isolation boundaries.

Confidence Score: 4/5

• Safe to merge with minor considerations - the implementation is solid but has extensive previous review comments that have been addressed or explained by the developer
• The code demonstrates thoughtful design with proper error handling, comprehensive testing, and addresses the core issue of parallel agent isolation. The extensive previous review thread shows the developer has been responsive to feedback and made deliberate architectural choices (e.g., symlinking nested dependencies for performance). The pattern matching logic is well-tested and the two-phase staging approach is robust. Score reflects solid implementation quality with no critical issues found.
• No files require special attention - all changes are well-structured and properly tested

Important Files Changed

Filename	Overview
cli/src/execution/sandbox.ts	Added centralized pattern matching functions (matchesPattern, isIgnored) with proper regex escaping and introduced DEFAULT_IGNORED list to filter infrastructure files during sandbox creation
cli/src/execution/parallel.ts	Refactored file filtering to use centralized DEFAULT_IGNORED patterns instead of hardcoded strings, with proper directory boundary checks and basename matching for recursive filtering
cli/src/execution/sandbox-git.ts	Implemented two-phase staging with safeGitAdd helper that batches files, filters ignored files via git check-ignore, and retries on lock contention for robust parallel execution

Sequence Diagram

sequenceDiagram
    participant Runner as Main Runner
    participant Agent as Parallel Agent
    participant Sandbox as Sandbox (sandbox.ts)
    participant Git as Git Operations (sandbox-git.ts)
    participant Filter as File Filter (parallel.ts)

    Runner->>Agent: Execute task in parallel
    Agent->>Sandbox: createSandbox(originalDir, sandboxDir)
    Sandbox->>Sandbox: Filter items using DEFAULT_IGNORED
    Sandbox->>Sandbox: isIgnored(item, DEFAULT_IGNORED, isDir)
    Sandbox->>Sandbox: Symlink dependencies (node_modules, .git)
    Sandbox->>Sandbox: Copy source files recursively
    Note over Sandbox: .ralphy/ directory excluded from sandbox
    Sandbox-->>Agent: Sandbox created with isolation
    
    Agent->>Agent: Execute task modifications
    Note over Agent: progress.txt instruction removed
    
    Agent->>Filter: getModifiedFiles(worktreeDir, workDir)
    Filter->>Filter: Compare file timestamps
    Filter-->>Agent: Return modified files list
    
    Agent->>Filter: Filter files using DEFAULT_IGNORED
    loop For each modified file
        Filter->>Filter: Check directory patterns (e.g., ".ralphy/")
        Filter->>Filter: Check basename patterns (e.g., "nul", "*.log")
        Filter->>Filter: matchesPattern(baseName, pattern)
    end
    Note over Filter: Infrastructure files filtered out
    Filter-->>Agent: Return filtered files
    
    Agent->>Git: commitSandboxChanges(filteredFiles)
    Git->>Git: Create new branch
    Git->>Git: Copy filtered files to original dir
    
    Note over Git: Phase 1: Stage explicitly filtered files
    Git->>Git: safeGitAdd(git, filesToStage)
    Git->>Git: Batch files (20 per batch)
    Git->>Git: Filter via git check-ignore
    Git->>Git: Retry on index.lock contention
    
    Note over Git: Phase 2: Catch additional files
    Git->>Git: git status (find untracked/modified)
    loop For each untracked file
        Git->>Git: Check against DEFAULT_IGNORED
        Git->>Git: Filter infrastructure files
    end
    Git->>Git: safeGitAdd(git, additionalFiles)
    Note over Git: Handles composer/npm installs in symlinked dirs
    
    Git->>Git: git commit
    Git->>Git: Return to original branch
    Git-->>Agent: Commit successful
    
    Agent-->>Runner: Task completed with branch

Loading

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 3 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 4 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{5 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Additional Comments (1)

cli/src/execution/sandbox-git.ts
logging modifiedFiles.length instead of filesToAdd.length - reports incorrect count when some files were filtered by .gitignore

		logDebug(`Agent ${agentNum}: Committed ${filesToAdd.length} files to ${branchName}`);

Prompt To Fix With AI

This is a comment left during a code review.
Path: cli/src/execution/sandbox-git.ts
Line: 127:127

Comment:
logging `modifiedFiles.length` instead of `filesToAdd.length` - reports incorrect count when some files were filtered by `.gitignore`

```suggestion
		logDebug(`Agent ${agentNum}: Committed ${filesToAdd.length} files to ${branchName}`);
```

How can I resolve this? If you propose a fix, please make it concise.

[greptile-apps]

_{2 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{5 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[BasselBlal]

I really appreciate the project and hope this gets accepted!
P.S. This is my first open source PR 🚀

[greptile-apps]

_{2 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}