The Agent Skills Directory

[COMMAND_EXECUTION]: The skill employs 'Bang-substitution' (!command) syntax in references/ticket-sources.md to execute the gh (GitHub) CLI tool on the host system at load or prep time. This mechanism runs shell commands locally to fetch ticket data.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from external sources (GitHub issues, Linear tickets) which is then interpolated into the agent's implement prompt.
Ingestion points: Data is fetched from GitHub Issues via gh issue list and Linear via API/MCP (described in references/ticket-sources.md).
Boundary markers: The skill lacks explicit boundary markers or 'ignore embedded instructions' warnings when inlining ticket titles, bodies, and comments into the implement prompt.
Capability inventory: The sandboxed agents have the capability to write code, execute tests, and perform reviews. The host-side orchestrator has the ability to push git branches and open pull requests.
Sanitization: There is no evidence of sanitization or filtering of the external ticket content before it is processed by the AI agents.
[EXTERNAL_DOWNLOADS]: The skill relies on the @ai-hero/sandcastle package and instructs the user to run npx sandcastle init and pnpm sandcastle, which involves downloading and executing code from external registries.
[CREDENTIALS_UNSAFE]: The workflow requires the use of sensitive tokens (GH_TOKEN, CLAUDE_CODE_OAUTH_TOKEN). The instructions suggest storing these in a .env file, which is a standard practice but ensures the skill has access to high-privilege credentials.

sandcastle