The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements an orchestrator role that manages data flow between multiple subagents and the user. This architecture creates a surface where data from one specialist could potentially influence the orchestrator's decisions or the prompts sent to other subagents.\n
Ingestion points: The orchestrator ingests untrusted data from specialist result objects, user-supplied requirement convergence records, and project artifacts such as the PRD and Design Doc (documented in SKILL.md and references/monorepo-flow.md).\n
Boundary markers: The skill utilizes "canonical serialization" and "active workflow input contracts" to structure data during handoffs between agents, though it does not specify rigorous escaping or sanitization protocols for all interpolated content.\n
Capability inventory: The orchestrator can invoke other subagents as fresh tool calls and execute shell commands including git commit and git status via Bash (SKILL.md).\n
Sanitization: Boundary management relies on structured JSON handoff contracts (HC-01 through HC-06) and platform-level agent isolation, rather than explicit content filtering.\n- [COMMAND_EXECUTION]: The orchestration workflow involves the direct execution of shell commands to update the repository state.\n
Evidence: The "Task Execution Cycle" defined in SKILL.md explicitly instructs the agent to "execute git commit with Bash" to finalize task completion and record implementation changes.

subagents-orchestration-guide