The Agent Skills Directory

[COMMAND_EXECUTION]: The script orchestrates the execution of external CLI tools (OpenAI Codex and Anthropic Claude) using subprocess.Popen. It constructs command lines by appending user-provided prompts to base commands.
[COMMAND_EXECUTION]: Configures sub-agents to bypass security confirmations using flags like --ask-for-approval never and --dangerously-skip-permissions in scripts/parallel_vibe.py, removing human-in-the-loop safety checks.
[COMMAND_EXECUTION]: Supports a shell runner type in scripts/parallel_vibe.py that executes arbitrary command templates; while it uses shlex.quote, the template itself remains a surface for command injection if sourced from an untrusted plan.
[PROMPT_INJECTION]: High surface for indirect prompt injection by passing user prompts directly to sub-agents.
Ingestion points: File: scripts/parallel_vibe.py (via --prompt and --plan-file arguments).
Boundary markers: Uses 'soft guardrail' instructions in wrap_thread_prompt which are non-binding for the sub-agents.
Capability inventory: File: scripts/parallel_vibe.py (uses subprocess.Popen for execution and shutil for file operations).
Sanitization: Employs shlex.quote for shell templates, but lacks content filtering for malicious instructions directed at sub-agents.
[DATA_EXFILTRATION]: File: scripts/parallel_vibe.py. Duplicates the current directory contents into multiple sub-workspaces. While common folders are excluded, any sensitive files not in the default exclusion list are duplicated and processed by AI tools.
[DATA_EXFILTRATION]: File: scripts/parallel_vibe.py. The symlink_policy can be configured to 'keep', allowing potential exposure of files outside the source directory via symbolic links.

parallel-vibe