The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads external data from the specs/ directory (requirements and task definitions) and interpolates it into prompts for coder and reviewer subagents. This constitutes a surface for indirect prompt injection if the project specifications originate from untrusted sources.
Ingestion points: specs/{feature}/README.md, specs/{feature}/requirements.md, and specs/{feature}/tasks/*.md are read in SKILL.md (Steps 1, 3, 4, 6).
Boundary markers: The templates in references/coder-prompt-template.md and references/review-prompt-template.md use Markdown headers to separate feature context and task instructions, though they do not include explicit 'ignore embedded instructions' directives.
Capability inventory: The orchestrator and subagents have capabilities for file writing, shell command execution (linting, typechecking), and git operations.
Sanitization: There is no evidence of sanitization or filtering of the markdown content before it is processed by the subagents.
[COMMAND_EXECUTION]: The skill performs shell commands to verify implementation quality and manage source control.
Evidence: Executes git add and git commit to finalize waves of work.
Evidence: Instructs the review subagent to run project-specific verification commands such as pnpm lint and pnpm typecheck (SKILL.md, Step 6).
[DYNAMIC_EXECUTION]: The skill facilitates the creation and verification of new code through subagents. While this is the primary intent, it involves the dynamic generation and execution of code within the project environment.
Evidence: Coder subagents are tasked with creating and modifying files based on requirements (SKILL.md, Step 4).
Evidence: The review agent executes the newly implemented or modified code via project-standard linting and testing tools (SKILL.md, Step 9).

implement-feature