The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides various shell commands for local environment analysis and GitHub runner management.
Uses gh api and gh run list to query GitHub Actions status and metadata.
Uses docker run to execute performance benchmarks and architecture checks (node, ldd).
Uses shell scripting (case, uname -m) to handle multi-architecture deployments safely.
[EXTERNAL_DOWNLOADS]: The skill references and utilizes external software components.
Recommends the use of myoung34/github-runner:latest, a third-party Docker image widely used for self-hosted runners.
Mentions nektos/act as an alternative tool for local workflow execution.
Mentions gitleaks as an example for architecture-aware installation logic.
[DATA_EXPOSURE]: The skill handles sensitive GitHub Actions registration tokens.
Implementation includes security best practices: writing the token to a temporary file with chmod 600 permissions and immediately deleting it after use to prevent exposure in docker inspect or shell history.
[INDIRECT_PROMPT_INJECTION]: The skill identifies and warns about the risk of executing untrusted code via GitHub Actions.
Ingestion points: The runner processes code, dependencies, and scripts provided in GitHub Pull Requests (SKILL.md).
Boundary markers: The skill explicitly warns that "the risk is not who opens the PR, but what code that job executes," and suggests using manual triggers or tags to mitigate risks.
Capability inventory: The runner executes arbitrary shell commands defined in workflows, potentially with access to the host via docker.sock if configured by the user (SKILL.md).
Sanitization: No automated sanitization is performed, as the intended purpose is to run the provided workflow code; the skill relies on user-initiated process isolation.

self-hosted-github-runner