The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell scripts to perform its tasks.
SKILL.md instructs the agent to run bash .agents/skills/codex-write-pr-description/scripts/context.sh to gather repository data.
It also invokes bash .agents/skills/codex-pr/scripts/validate-description.sh for output validation, indicating a dependency on a related skill from the same environment.
The context.sh script uses standard binaries such as git and gh (GitHub CLI) to retrieve commit history, diffs, and issue data.
[INDIRECT_PROMPT_INJECTION]: The skill processes external, untrusted data that could contain malicious instructions intended to influence the AI's behavior.
Ingestion points: The skill reads commit messages, code diffs, existing pull request bodies, and issue titles through the context.sh script.
Boundary markers: While context.sh uses section headers (e.g., == diff ==), it does not provide explicit instructions or delimiters to the LLM to prevent it from following commands embedded within those data sources.
Capability inventory: The skill allows the agent to write files (PR_AGENTDESC.md) and execute subsequent shell commands for validation.
Sanitization: The scripts do not perform any sanitization, filtering, or escaping of the content retrieved from the repository history before passing it to the agent's context.

codex-write-pr-description