claude-code-action_skill

Overview

This skill provides a knowledge base and practical guidance for creating and configuring anthropics/claude-code-action GitHub workflows. It focuses on authentication options, common workflow patterns (interactive, CI auto-fix, nightly), permission settings, tool allowlisting, and cost/security controls. Follow the patterns and inputs here to safely automate Claude-powered code tasks in CI and PR contexts.

How this skill works

The skill inspects and documents workflow inputs, triggers, and best-practice blocks you should include in GitHub Actions using claude-code-action. It explains the authentication choices (OAuth, API key, AWS Bedrock, GCP Vertex), standard permissions, allowed tools, and recommended guards against infinite loops and excessive cost. It also includes reusable snippets for interactive, CI auto-fix, and scheduled workflows and guidance for passing MCP config and secrets.

When to use it

• Add interactive responses to PRs or issue mentions when contributors tag @claude.
• Run automated code fixes after failing CI workflows to propose or open fix PRs.
• Perform maintenance tasks nightly or on a schedule (test improvements, coverage).
• Integrate Claude into CI for linting, refactors, or repository-wide updates.
• Set up secure, cloud-hosted Claude access via AWS Bedrock or GCP Vertex.

Best practices

• Prefer CLAUDE_CODE_OAUTH_TOKEN (requires Claude Pro/Max) for most CI setups; store as a GitHub secret.
• Never hardcode credentials; reference secrets like ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}.
• Use permissions block with minimal required scopes (contents: write, pull-requests: write, issues: write, actions: read, id-token: write).
• Guard CI automations to prevent infinite loops and exclude main/stable branches from auto-fix runs.
• Limit API usage with --max-turns per role: interactive unlimited, CI auto-fix ~25, nightly ~30.
• Allowlist only necessary tools (file ops, git, package managers, gh) and sanitize prompt inputs to avoid injection.

Example use cases

• Trigger Claude on PR comment to generate a concise code change suggestion or unit tests.
• Automatically run a CI quality-check workflow; on failure, have Claude propose fixes in a new branch and PR.
• Run nightly test-improvement passes that update flaky tests or improve coverage with an opt-in variable.
• Use Claude behind AWS Bedrock or GCP Vertex for enterprise-hosted model inference with cloud credentials.
• Integrate MCP configuration via claude_args to route requests through a private control plane.