The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python adapter scripts to execute external command-line tools including git, claude, and codex. This execution is the core functionality of the skill, used to manage code repositories and delegate tasks to sub-agents. The implementation includes significant security controls:
Commands are invoked with validated arguments and specific environment filtering.
The shutil.which utility is used to resolve binary paths safely.
The subprocess.run calls are constrained to specific, necessary operations like authentication status checks and version reporting.
[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection because it processes output from external AI models (workers) which could contain malicious instructions. However, the skill is specifically designed to mitigate this risk through several defensive layers:
Ingestion points: The adapter scripts in scripts/claude_delegate.py and scripts/codex_delegate.py read and parse the stdout/stderr and JSON results returned by worker CLIs.
Boundary markers: scripts/delegate_common.py implements a hardened prompt envelope using random 16-byte hex boundary tokens (<<<mission {boundary}>>>) and explicit instructions to the worker to treat all input as untrusted data.
Capability inventory: The skill possesses the ability to write files (receipts/artifacts) and execute shell commands via git and model CLIs.
Sanitization: All filesystem paths are validated against a strict regex ([A-Za-z0-9][A-Za-z0-9._-]*) that prevents absolute paths, parent directory traversal (..), or hidden files. It also explicitly checks for and blocks symlink traversal.

vibe-orchestrate