The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The orchestrator is designed to ingest and act upon data generated by sub-agents, which constitutes a significant attack surface for indirect prompt injection.
Ingestion points: The Coordinator (the agent using this skill) consumes pending messages and status reports from sub-agents using orca orchestration check and orca orchestration worker-read (SKILL.md, Section 6).
Boundary markers: The protocol lacks explicit boundary markers or instructions to treat sub-agent output as untrusted content, which could lead the Coordinator to follow instructions embedded in worker reports.
Capability inventory: The skill utilizes powerful capabilities across various scripts and instructions, including git branch manipulation (git merge, git rebase), build/test execution (npm test, npm run build), and lifecycle management for other agents (orca orchestration worker-start, orca terminal close).
Sanitization: No sanitization, escaping, or validation of the content received from sub-agents is performed before the Coordinator processes the information or presents it to the user.
[COMMAND_EXECUTION]: The orchestration workflow involves the automated execution of shell commands to manage codebases and run validation suites.
Evidence: The skill instructs the agent to perform rebase operations (git -C ... rebase), merges (git merge --no-ff), and to execute user-defined or environment-defined commands such as ${VALIDATION_TEST_CMD:-npm test} and ${VALIDATION_DEPLOY_CMD} (SKILL.md, Section 6).

agent-orchestration