The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill mirrors code from the 'pulumi/autoregistry' monorepo into public repositories. It clones this repository and executes build-time operations on its contents, including 'go build', 'make', and 'npm install'. This establishes a surface for indirect prompt injection if the source repository were to contain malicious instructions or code.
Ingestion points: 'git clone' and 'git archive' operations in 'scripts/export.py' and 'scripts/verify.py'.
Boundary markers: Absent.
Capability inventory: Extensive use of 'subprocess.run', 'make' targets, 'go' toolchain, and 'npm' in 'scripts/verify.py' and 'scripts/examples_gate.py'.
Sanitization: Absent; the skill is designed to mirror and rewrite code content rather than sanitize natural language instructions.
[DYNAMIC_EXECUTION]: The 'scripts/examples_gate.py' script utilizes the Python 'compile()' function to perform syntax validation on Python programs fetched from the monorepo. It also executes a dedicated 'RESOLVER' script via a child process that uses 'importlib.import_module()' to verify imports in the mirrored code. This represents dynamic execution of content derived from an external source as part of its primary verification purpose.
[COMMAND_EXECUTION]: The skill makes heavy use of 'subprocess.run()' to manage its environment, install the toolchain, and execute the mirroring process. Notable patterns include 'eval "$(python3 scripts/bootstrap.py)"' in 'SKILL.md' to configure shell environment variables and various 'gh' and 'git' commands for repository management and permissions probing.
[EXTERNAL_DOWNLOADS]: 'scripts/bootstrap.py' downloads and installs binaries for Go, Pulumi, Node.js, the .NET SDK, and the Java Development Kit. These downloads are performed from well-known official domains (e.g., 'go.dev', 'github.com', 'microsoft.com', 'nodejs.org') and are rigorously verified against hardcoded SHA-256 or SHA-512 hashes within the script.

pulumi-labs-export