The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes parallel child processes to perform research subtasks using a custom Python runner (scripts/run_children.py).
It utilizes the subprocess.run function to call the platform's codex exec CLI tool.
Security is maintained by explicitly enabling platform-level sandboxing (e.g., --sandbox workspace-write and --ephemeral) for all generated subtasks.
Command arguments are constructed as lists and executed without a shell, preventing shell injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external research sources, creating a potential surface for indirect prompt injection where untrusted data could influence agent instructions.
Ingestion points: External websites, documents, and datasets analyzed during the subtask execution phase by child agents.
Boundary markers: The skill instructs sub-agents to structure their output as Markdown with cited sources, although it does not implement unique random delimiters for raw data interpolation.
Capability inventory: The skill performs localized file writes within a .research/ directory and manages parallel process execution.
Sanitization: The implementation includes a robust resolve_within utility that validates all file paths against the intended run directory to prevent directory traversal attacks. Furthermore, child processes are restricted to a workspace-write sandbox to limit their impact.

deep-research