The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes an included Python script (scripts/packs-resolve.py) during the fixing phase to resolve project-specific rules. The script is called via a shell wrapper to ensure compatibility with available Python interpreters.
[EXTERNAL_DOWNLOADS]: The skill fetches 'Compound Packs' from remote Git repositories as specified in the local .compound-engineering/config.yaml. The implementation includes security features such as using --end-of-options to prevent command injection and strictly verifying cache directory ownership and permissions (0700).
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from issue trackers (e.g., GitHub, Linear, Jira) and external rule files (Packs), which create a surface for indirect prompt injection. The skill includes a mandatory evidence chain for its analysis and explicitly instructs the agent to disregard instructions embedded in these data sources, effectively mitigating the risk of obedience to injected commands.

ce-debug