The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted instructions from markdown files in the .scratch/ directory that define work items and acceptance criteria. This content is used to drive code changes and command execution without isolation. * Ingestion points: Step 1 in SKILL.md fetches ticket/spec bodies and comments from files in .scratch/ and tracker manifests. * Boundary markers: Absent. The instructions do not define delimiters or provide warnings to disregard embedded commands in the input data. * Capability inventory: Shell execution of git, GitHub CLI, and framework-specific commands (Laravel artisan), alongside code modification through integrated TDD and verification sub-skills. * Sanitization: Absent. The agent translates ticket criteria into code and tests without sanitizing the input against malicious instructions.
[COMMAND_EXECUTION]: The skill triggers shell commands throughout its workflow for repository management and testing. Evidence includes Git operations (status, commit, checkout) in SKILL.md and framework-specific utility and test commands (php artisan test, vendor/bin/pint) listed in references/laravel.md.
[DATA_EXFILTRATION]: The skill performs sensitive file access and remote operations as part of its development lifecycle. Evidence: Step 2 of the PR delivery section in references/spec-run.md instructs the agent to copy local environment configurations (.env files) when bootstrapping worktrees. Additionally, it uses the GitHub CLI (gh pr create) to transmit code and verification evidence to remote repositories.

implement