The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/route.mjs executes several system commands to manage the background Codex worker and its environment:
bash -c is used to launch the codex CLI in a detached background process with redirected input and output.
node is used to execute the codex-companion.mjs script from the Codex plugin.
pkill -f is used to terminate the codex app-server process to resolve stale authentication states.
which or where is used to locate the codex binary on the system path.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a proxy for a sub-agent (Codex), creating a surface for indirect prompt injection:
Ingestion points: Untrusted data enters via the command-line prompt argument or the contents of a file provided via the --prompt-file flag in scripts/route.mjs.
Boundary markers: The wrapper does not inject explicit boundary markers or safety instructions into the prompt; it relies on the user to scope the task correctly.
Capability inventory: The skill can write to the file system (logs and metadata in ~/.claude/route-jobs/) and execute subprocesses (via spawn and spawnSync in scripts/route.mjs).
Sanitization: The wrapper uses shell-safe argument passing ("$@") when invoking the background worker to prevent command injection into its own shell execution environment.

route