The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Codex CLI (codex exec) and standard shell commands to verify environment readiness, manage git worktrees, and coordinate subagent execution across multiple repositories.
[DATA_EXFILTRATION]: Security investigation reports are transmitted to an external service (Cosense/Scrapbox) as defined in the skill's configuration. This behavior is documented and represents the primary intended functionality for reporting and triaging issues.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted source code from repositories during auditing, which presents a surface for indirect prompt injection attacks.
Ingestion points: Source code from multiple target repositories is read for analysis.
Boundary markers: The skill explicitly manages a checklist.md file and instructs the agent to remove assertions or severities that violate the 'responsibility boundary' between setup and reporting stages.
Capability inventory: The skill can perform file writes, git operations (including merges and branch management), and execute sub-agents for automated code fixing.
Sanitization: Mitigation is achieved by requiring high-tier models (Opus tier or flagship Codex models) and utilizing structural verification of report outputs.
[EXTERNAL_DOWNLOADS]: Instructs users to install the @helpfeel/cosense-cli package via the official NPM registry to facilitate report management with the well-known service used by the tool.

codepatrol