The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents a pattern for ingesting data from project-local files (.Codex/*.local.md) using YAML frontmatter and markdown. This pattern allows external data to influence agent behavior and shell script logic (hooks), which is a common attack vector for indirect prompt injection.\n
Ingestion points: Files matching the pattern .Codex/plugin-name.local.md are read by agents and hooks as described in SKILL.md.\n
Boundary markers: The pattern utilizes YAML frontmatter delimiters (---) to separate configuration from content, as detailed in references/parsing-techniques.md.\n
Capability inventory: The skill provides examples of using this ingested data in shell scripts (e.g., examples/read-settings-hook.sh) and for notifying other sessions via tmux send-keys in the multi-agent-swarm example in references/real-world-examples.md.\n
Sanitization: The skill includes explicit guidance on sanitizing user input, validating numeric ranges, and checking for path traversal or sensitive file patterns (e.g., .env, secret) in SKILL.md and the examples/read-settings-hook.sh template.

plugin-settings