The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides a wrapper for executing caller-supplied commands through the codex_exec_guarded function. It captures process output and manages exit statuses.
Evidence: The skill description states it runs "exactly one caller-supplied Codex prompt" and captures the result.
[DYNAMIC_EXECUTION]: The skill executes code that is provided at runtime via environment variables or stdin piping. It sources local libraries to facilitate this execution.
Evidence: The script sources . "$AGENTOPS_ROOT/scripts/lib/codex-exec.sh" and uses the CODEX_EXEC_PROMPT_ARG variable for execution.
[INDIRECT_PROMPT_INJECTION]: The skill serves as a high-capability sink for processing data generated by other components, which may contain instructions targeting the execution environment.
Ingestion points: Accepts a "caller-supplied prompt" via the PROMPT argument or as piped input to the process.
Boundary markers: The skill documentation does not define specific delimiters for untrusted content within the prompt, relying instead on environment-level sandboxing.
Capability inventory: The skill can execute shell processes with potential workspace-write and network access depending on the sandbox configuration (CODEX_EXEC_SANDBOX).
Sanitization: There is no explicit sanitization or filtering of the prompt content; security relies on the read-only sandbox recommendation and the capture cap (CODEX_EXEC_MAX_OUTPUT_BYTES).

codex-exec