## Problem
AI assistants were consistently connecting SplitInBatches node outputs backwards because:
- Output index 0 = "done" (runs after loop completes)
- Output index 1 = "loop" (processes items inside loop)
This counterintuitive ordering caused incorrect workflow connections.
## Solution
Enhanced the n8n-mcp system to expose and clarify output information:
### Database & Schema
- Added `outputs` and `output_names` columns to nodes table
- Updated NodeRepository to store/retrieve output information
### Node Parsing
- Enhanced NodeParser to extract outputs and outputNames from nodes
- Properly handles versioned nodes like SplitInBatchesV3
### MCP Server
- Modified getNodeInfo to return detailed output descriptions
- Added connection guidance for each output
- Special handling for loop nodes (SplitInBatches, IF, Switch)
### Documentation
- Enhanced DocsMapper to inject critical output guidance
- Added warnings about counterintuitive output ordering
- Provides correct connection patterns for loop nodes
### Workflow Validation
- Added validateSplitInBatchesConnection method
- Detects reversed connections and provides specific errors
- Added checkForLoopBack with depth limit to prevent stack overflow
- Smart heuristics to identify likely connection mistakes
## Testing
- Created comprehensive test suite (81 tests)
- Unit tests for all modified components
- Edge case handling for malformed data
- Performance testing with large workflows
## Impact
AI assistants will now:
- See explicit output indices and names (e.g., "Output 0: done")
- Receive clear connection guidance
- Get validation errors when connections are reversed
- Have enhanced documentation explaining the correct pattern
Fixes#97🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Create benchmark test suites for critical operations:
- Node loading performance
- Database query performance
- Search operations performance
- Validation performance
- MCP tool execution performance
- Add GitHub Actions workflow for benchmark tracking:
- Runs on push to main and PRs
- Uses github-action-benchmark for historical tracking
- Comments on PRs with performance results
- Alerts on >10% performance regressions
- Stores results in GitHub Pages
- Create benchmark infrastructure:
- Custom Vitest benchmark configuration
- JSON reporter for CI results
- Result formatter for github-action-benchmark
- Performance threshold documentation
- Add supporting utilities:
- SQLiteStorageService for benchmark database setup
- MCPEngine wrapper for testing MCP tools
- Test factories for generating benchmark data
- Enhanced NodeRepository with benchmark methods
- Document benchmark system:
- Comprehensive benchmark guide in docs/BENCHMARKS.md
- Performance thresholds in .github/BENCHMARK_THRESHOLDS.md
- README for benchmarks directory
- Integration with existing test suite
The benchmark system will help monitor performance over time and catch regressions before they reach production.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Enhanced database adapter to support multiple WASM file resolution strategies
- Added require.resolve() for reliable package location in npm environments
- Made better-sqlite3 an optional dependency
- Improved error handling with clear messages
- Updated version to 2.7.13
- Updated CHANGELOG and README badges
- Added runtime FTS5 detection in database adapters
- Removed FTS5 from required schema to prevent "no such module" errors
- FTS5 tables/triggers created conditionally only if supported
- Template search automatically falls back to LIKE when FTS5 unavailable
- Works in ALL SQLite environments (Claude Desktop, restricted envs, etc.)
This ensures search_templates() works correctly regardless of SQLite build,
while still providing optimal performance when FTS5 is available.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Fixed issue where Docker images using sql.js adapter returned boolean fields
as strings, causing is_trigger=0 to evaluate as true instead of false.
Changes:
- Added convertIntegerColumns() to sql.js adapter to convert SQLite integers
- Updated server.ts and node-repository.ts to use Number() conversion as backup
- Added test script to verify fix works with sql.js adapter
This fixes webhook, cron, interval, and emailReadImap nodes showing
isTrigger: false in Docker deployments.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- SQLite doesn't allow datetime('now') in CHECK constraints
- Drop tables before recreating to ensure clean schema
- 6-month filtering is already handled in application logic
- Successfully fetched and stored 199 templates
🤖 Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
- Add 4 new MCP tools for workflow templates
- Integrate with n8n.io API to fetch community templates
- Filter templates to last 6 months only
- Store templates in SQLite with full workflow JSON
- Manual fetch system (not part of regular rebuild)
- Support search by nodes, keywords, and task categories
- Add fetch:templates and test:templates npm scripts
- Update to v2.4.1
🤖 Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
- Add InitializeRequestSchema handler to MCP server
- Implement stdout flushing for Docker environments
- Create stdio-wrapper for clean JSON-RPC communication
- Update docker-entrypoint.sh to prevent stdout pollution
- Fix logger to check MCP_MODE before level check
These changes ensure the MCP server responds to initialization requests
within Claude Desktop's 60-second timeout when running in Docker.
- Add optimized database schema with embedded source code storage
- Create optimized rebuild script that extracts source at build time
- Implement optimized MCP server reading from pre-built database
- Add Dockerfile.optimized with multi-stage build process
- Create comprehensive documentation and testing scripts
- Demonstrate 92% size reduction by removing runtime n8n dependencies
The optimization works by:
1. Building complete database at Docker build time
2. Extracting all node source code into the database
3. Creating minimal runtime image without n8n packages
4. Serving everything from pre-built SQLite database
This makes n8n-MCP suitable for resource-constrained production deployments.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
