claude-code-gitea-action/PERFORMANCE-ANALYSIS.md

Claude Code Gitea Action Performance Analysis

Executive Summary

The Claude Code Gitea Action was experiencing significant performance issues, with total execution times of ~20 minutes for simple tasks that should complete in 2-3 minutes. This analysis identifies the root causes and provides solutions that can reduce execution time to 5-7 minutes.

Performance Breakdown

Current State (Before Optimizations)

• Total execution time: ~20 minutes
• Claude AI Assistant step: 10m51s
  • Actual Claude execution: ~2.5 minutes
  • Setup/installation: ~2 minutes
  • MCP server shutdown wait: ~6 minutes
• Job cleanup: 7-9 minutes

Root Causes Identified

1. MCP Server Shutdown Issues (6 minutes delay)

The MCP servers (gitea-mcp-server.ts and local-git-ops-server.ts) were not shutting down properly after Claude completed execution. They use StdioServerTransport which keeps stdin/stdout connections open, waiting for an EOF signal that never arrived.

Impact: 6 minutes of unnecessary waiting

2. Container Layering

The workflow uses a nested container setup:

runs-on: ubuntu-latest
container:
  image: node:18-bullseye

This creates multiple container layers that significantly increase cleanup time.

Impact: 5-6 minutes of additional cleanup time

3. Failed Cache Attempts

Multiple components attempt to use caching but fail with timeout errors:

• Bun package cache
• Node.js cache
• npm cache

Impact: 10-20 seconds per cache attempt, plus error noise in logs

4. Duplicate Package Installation

The action installs packages twice:

• First bun install for action dependencies
• Second bun install for claude-code package

Impact: ~3-4 seconds (minor but unnecessary)

Implemented Solutions

1. Fixed MCP Server Shutdown

Added proper shutdown handling to both MCP servers:

• Listen for stdin EOF signal
• Properly close transport connections
• Add signal handlers for SIGTERM, SIGINT, SIGHUP
• 5-minute timeout as safety net

Expected savings: 6 minutes

2. Disabled Caching

Added cache: false to Bun setup to avoid timeout errors.

Expected savings: 10-20 seconds, cleaner logs

3. Created Optimized Workflow

Provided claude-optimized.yml that removes the container layer.

Expected savings: 5-6 minutes in cleanup time

4. Custom Container Option

Created Dockerfile.runner for a pre-built container with all dependencies.

Expected savings: 1-2 minutes in setup time

Expected Results

With Current Workflow + MCP Fixes

• Total time: ~11-13 minutes (down from 20)
• Savings: ~7-9 minutes (35-45% improvement)

With Optimized Workflow (No Container Layer)

• Total time: ~6-7 minutes (down from 20)
• Savings: ~13-14 minutes (65-70% improvement)

With Custom Pre-built Container

• Total time: ~5-6 minutes (down from 20)
• Savings: ~14-15 minutes (70-75% improvement)

Recommendations

Immediate Actions (No Infrastructure Changes)

1. Deploy the MCP server fixes - This alone saves 6 minutes
2. Monitor the logs to confirm MCP servers shut down immediately after Claude completes

Short-term Optimizations

1. Switch to the optimized workflow (remove container layer) for additional 5-6 minute savings
2. Test with different runner configurations if available in your Gitea setup

Long-term Optimizations

1. Build and use custom container image with pre-installed dependencies
2. Consider runner-level optimizations:
   • Use SSD storage for faster Docker operations
   • Allocate more resources to runners
   • Use local Docker registry for faster image pulls
3. Optimize the base action to reduce duplicate installations

Additional Findings

Token Permissions

The action requires specific token scopes:

• write:repository - For code changes (already working)
• write:issue - For commenting on issues
• write:pull_request - For creating PRs

Ensure your RUNNER_ACCESS_TOKEN has all required scopes.

OAuth Authentication

OAuth authentication via claude_credentials is working correctly after:

• Extracting access token from credentials JSON
• Passing via claude_code_oauth_token parameter
• Setting CLAUDE_CODE_OAUTH_TOKEN environment variable

Monitoring

After implementing these changes, monitor:

1. MCP server shutdown logs - Should show "Stdin closed, shutting down..."
2. Total step execution times in Gitea UI
3. Any remaining timeout or connection errors

Conclusion

The performance issues were primarily caused by improper MCP server shutdown handling and unnecessary container layering. The implemented fixes should reduce execution time by 65-75%, bringing a 20-minute workflow down to 5-7 minutes.