The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates a workflow where the agent processes code and reports generated by an external implementer (OpenAI Codex). This creates a vulnerability surface for instructions embedded in the implementer's output to potentially influence the orchestrator agent's subsequent actions.
Ingestion points: SKILL.md and references/dispatch-and-poll.md describe reading result.json, particularly the finalMessage and touchedFiles fields.
Boundary markers: The skill advises using structured XML tags in the task brief and defining a clear output contract to delimit implementer responses.
Capability inventory: The orchestrator agent is expected to execute shell commands for project verification (testing, linting) and git operations.
Sanitization: The skill implements robust defense-in-depth by instructing the agent to never trust the implementer's self-report, requiring manual diff verification and independent execution of project gate commands.
[EXTERNAL_DOWNLOADS]: The skill documentation references external tools and packages from well-known sources.
Evidence: Mentions installation of the @openai/codex package via npm. OpenAI is recognized as a well-known service provider.
[COMMAND_EXECUTION]: The skill provides instructions for the orchestrator agent to execute shell commands to dispatch tasks.
Evidence: SKILL.md contains example commands for running a relay script using node and project-specific "gate" commands (tests, build scripts).
[SAFE]: The skill explicitly omits executable scripts (specifically scripts/relay.mjs) from its distribution. It provides documentation-only templates and advises the user to separately obtain and audit the necessary runtime components, reducing the risk of executing unverified bundled code.

codex-delegate