The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute commands defined within a repository-root file named ORCHESTRATION.md. These "lane-admission preflights" allow the repository to specify toolchain checks, resource constraints, or environmental setup routines that the agent must run before initiating work. While this allows for arbitrary command execution based on repository content, the skill includes instructions to validate the file and stop execution if it is malformed or contradictory.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests untrusted data from external sources, including GitHub issue descriptions, roadmap items, and pull request comments, to generate task packets for subagents.
Ingestion points: Specifically reads ORCHESTRATION.md, roadmap items for the /implement-idea tool, and existing GitHub issues/PRs for task context.
Boundary markers: The instructions mandate the use of isolated context packets for subagents and require validation of the orchestration reference before execution.
Capability inventory: The skill has broad capabilities including git operations (commit, push, merge), milestone management, and the ability to spawn subagents for implementation and code review.
Sanitization: The skill performs validation of the orchestration policy and uses specific "Definitions of Ready and Done" to gate the intake of new work.
[DATA_EXPOSURE]: The skill manages a persistent event ledger and handoff state within a local .agent/ directory. It accesses authenticated GitHub credentials and environment state (worktrees) to perform its coordination tasks, but there is no evidence of exfiltration to third-party domains.

milestone-rush