The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches external content such as issue bodies, comments, and pull request diffs from remote trackers like GitHub and GitLab based on user input (Stage 0
Triage).
[COMMAND_EXECUTION]: Uses platform-specific CLI tools (gh for GitHub, glab for GitLab) to perform operations like creating issues, linking sub-issues, and managing dependencies on the configured tracker.
[REMOTE_CODE_EXECUTION]: Includes specific procedures for verifying external pull requests that may involve code execution. This is mitigated by a mandatory safety protocol requiring static review first, followed by execution only in a disposable security sandbox (no secrets, no write credentials, network denied) and requiring explicit, separate user approval for commands.
[PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data (issue reports, PR descriptions, and comments), creating a surface for indirect prompt injection.
Ingestion points: Stage 0 Triage reads external issue/PR content and metadata.
Boundary markers: The skill uses an 'Agent Brief' as the authoritative contract for implementation, framing the original untrusted content as context only.
Capability inventory: File system writes (.agents/, docs/), tracker CLI interaction (gh, glab), and sandboxed shell execution.
Sanitization: No specific sanitization or delimiters are mandated for external text ingestion.

vibe-plan