The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from GitHub issue bodies and comments via the gh issue view command. This data is used to generate 'briefs' and instructions for subagents, creating a surface for indirect prompt injection attacks where an attacker could place malicious instructions in a comment to influence agent behavior.
Ingestion points: GitHub issue content fetched via gh issue view and references to ticket-evidence.md.
Boundary markers: Utilizes a durable 'ledger' and 'source map' to maintain state, alongside 'locked answers' derived from user interaction.
Capability inventory: The skill possesses significant capabilities including file writing, git operations (commit/push), and GitHub API mutations (PATCHing comments, updating PRs).
Sanitization: Employs several guardrails such as the preflight-mutations tool for verifying state before writes, harden-plan for plan validation, and a 'Grill the user' phase to establish authoritative requirements before autonomous work begins.
[COMMAND_EXECUTION]: The skill executes various system and development tools as part of its orchestration and implementation phases.
Evidence: Invokes gh (GitHub CLI) for repository interaction, git for source control management, and specialized tools like browser-qa for UI testing and type-check for code validation.
[EXTERNAL_DOWNLOADS]: Fetches data and media from GitHub's infrastructure.
Evidence: Downloads issue metadata, comment history, and attached images. These operations target GitHub, a well-known service, and include authentication and error-handling checks.
[PERSISTENCE]: Maintains state across sessions using local files, ensuring continuity of the 'unattended' run.
Evidence: Stores progress, commit SHAs, and next actions in a ledger (e.g., docs/<ticket>-progress.md) and task lists (tasks/todo.md). This allows the agent to resume complex workflows even after process restarts or memory compaction.

executing-tickets-with-subagents