The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The orchestrator fetches sub-issue metadata (numbers and titles) from GitHub and interpolates them into prompt templates for child agents, creating an attack surface for malicious issue titles.
Ingestion points: Sub-issue data is fetched in SKILL.md (Step 1) using gh api graphql and stored for use in prompts.
Boundary markers: The prompt skeleton in package-prompt.md lacks explicit boundary markers or instructions to treat the $TITLE variable as untrusted data.
Capability inventory: The skill possesses significant capabilities, including git branch management, gh PR merging, and execution of local shell scripts and package managers.
Sanitization: No evidence of sanitization or character escaping for the fetched issue titles was found.
[COMMAND_EXECUTION]: The skill performs extensive shell command execution and calls local scripts to manage the development lifecycle.
Evidence: SKILL.md executes node .agents/skills/verify/scripts/default-branch.mjs to resolve branching logic and gh commands for repository interaction.
Evidence: package-prompt.md instructs agents to execute scripts/worktree-up.sh and various git and npm commands.
Evidence: ops.md references the use of vercel inspect and npm run verify for deployment diagnostics.

orchestrate