The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several shell commands to manage the development lifecycle, including pnpm install for dependency management, gw ci --affected --fail-fast --diagnose for continuous integration checks, and uv run to execute the local glimpse tool for visual verification.
[EXTERNAL_DOWNLOADS]: The skill performs external downloads via pnpm install, which fetches packages from the Node.js registry to build the requested feature.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests a user-provided feature specification and passes it to multiple subagents with broad file-system and tool-execution permissions.
Ingestion points: The feature specification is received in Phase 1 of SKILL.md and interpolated into subagent templates in references/conductor-dispatch.md.
Boundary markers: The templates use simple curly-brace interpolation (e.g., {feature_spec}) without explicit instructions to disregard potentially malicious commands embedded in the specification.
Capability inventory: Subagents (Elephant, Turtle, etc.) have the ability to modify project files, while the conductor agent executes shell commands (pnpm, gw, uv).
Sanitization: No explicit input sanitization or verification logic is present to filter the user-provided specification before it is processed by the subagents.

gathering-feature