The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The skill uses the dynamic context injection syntax to run configuration discovery scripts at load time. This is used to resolve paths and environment variables necessary for the skill's operation. Evidence: !${CLAUDE_PLUGIN_ROOT}/scripts/discover-config.sh in SKILL.md.\n- [COMMAND_EXECUTION]: Extensive use of shell commands is employed to manage the development lifecycle, including repository synchronization, branch management, and artifact cleanup. Evidence: The skill uses git checkout, git pull, git worktree, gh pr view, yq, and rm across various phases for repository and PR management.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from various external sources including issue trackers and project files, which could theoretically contain malicious instructions meant to influence the agent. \n
Ingestion points: SKILL.md references loading items from backends (GitHub/Trello), reading project context files, domain terminology files, and grepping backlog Markdown files.\n
Boundary markers: The skill incorporates mandatory user approval phases ("Never auto-build") and uses structured YAML for tool delegation, which serves as a functional boundary for agent actions.\n
Capability inventory: The agent has permissions to modify local repository files, interact with the GitHub CLI, and delegate tasks to an external execution tool.\n
Sanitization: No explicit sanitization of text content from issue trackers is performed before being processed or passed to downstream tools.\n- [REMOTE_CODE_EXECUTION]: The skill delegates implementation and review tasks to an external harness tool. While this is the intended functionality, it involves transmitting project context and code for remote execution. Evidence: Phases 2B and 2C in SKILL.md invoke harness:delegate with operation: execute and operation: review.

sprint-dev