The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on multiple Bash scripts (setup.sh, session-start.sh, prompt-inbox.sh, stop-inbox.sh) to perform installation, manage local state in ~/.relay, and interact with the Relaycast API.
[EXTERNAL_DOWNLOADS]: The MCP server configuration in codex-config/config.toml uses npx -y agent-relay mcp to fetch and execute the agent-relay package from the public NPM registry at runtime.
[INDIRECT_PROMPT_INJECTION]: The hooks/prompt-inbox.sh script periodically polls an external API and injects received messages into the UserPromptSubmit hook context, which could allow external actors to influence agent behavior.
Ingestion points: External messages are fetched via curl from api.relaycast.dev in hooks/prompt-inbox.sh.
Boundary markers: Messages are injected into the context with a text header but lack robust delimiters or specific instructions to the model to treat the content as untrusted data.
Capability inventory: The skill provides access to Relaycast MCP tools for messaging, agent registration, and worker spawning.
Sanitization: The hooks/prompt-inbox.sh script replaces newlines in message text with spaces, but no semantic filtering or instruction sanitization is performed.
[PERSISTENCE]: The scripts/setup.sh script automatically modifies the agent's .codex/hooks.json file to register scripts that run on every session start, prompt submission, and agent exit, ensuring a persistent presence in the agent's runtime environment.

agent-relay