The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided goals and task acceptance criteria which are subsequently used to construct prompts for a cloud-based execution service. This creates a surface for potential indirect prompt injection attacks.\n
Ingestion points: The goal and acceptance parameters ingested by the prepare_local_plan and build_codex_cloud_dispatch_batch functions in scripts/cursor_orchestrate_adapters.py.\n
Boundary markers: The skill structures the outbound prompt using explicit headers such as 'Goal:', 'Task:', 'Role:', and 'Acceptance:' to help the remote model distinguish between system-provided metadata and user data.\n
Capability inventory: The skill facilitates the construction of shell-free argument tuples for codex cloud exec, codex cloud status, codex cloud diff, and codex cloud apply commands.\n
Sanitization: The _text validation function in scripts/cursor_orchestrate_adapters.py proactively scans input for regex-defined credential patterns (including OpenAI keys, GitHub tokens, and AWS IDs) and blocks forbidden path markers like .ssh and .git.\n- [COMMAND_EXECUTION]: The skill dynamically generates command-line arguments for the codex utility. It enforces strict character validation for identifiers and environment variables using the _cli_identifier function to prevent command injection or flag manipulation.\n- [REMOTE_CODE_EXECUTION]: The primary function of the skill is to coordinate code execution on the remote 'Codex Cloud' platform. Security is managed through a policy of requiring explicit user authorization for high-stakes operations like task creation and applying code changes.

cursor-orchestrate