The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill configuration explicitly enables flags that bypass human-in-the-loop safety approvals for multiple agent tools.
In config/cli-config.yaml, the auto_approve_flag is set to --dangerously-skip-permissions for the claude and antigravity vendors.
Similar bypass flags like --yolo and --trust-all-tools are configured for the qwen, cursor, and kiro agents.
These settings allow sub-agents to perform potentially destructive actions or access sensitive resources without the user's explicit consent, effectively escalating their effective permissions within the system.
[COMMAND_EXECUTION]: The skill uses wrapper scripts to execute external command-line tools for agent spawning and verification.
scripts/spawn-agent.sh and scripts/parallel-run.sh execute the oma binary with user-supplied arguments, which are then passed to external LLM CLIs.
scripts/verify.sh executes verification logic for different agent types using the oma verify command.
[INDIRECT_PROMPT_INJECTION]: The orchestration workflow involves ingesting complex instructions and managing data flow between multiple specialized agents, creating a risk for malicious instructions to be propagated through the system.
Ingestion points: The orchestrator reads task requirements, project configurations, and sub-agent progress/result files (e.g., result-*.md described in SKILL.md and resources/memory-schema.md).
Boundary markers: The skill uses structured markdown templates and a centralized task-board.md to manage state, but it lacks explicit instruction-ignore delimiters in the sub-agent prompt templates defined in resources/subagent-prompt-template.md.
Capability inventory: The skill possesses extensive capabilities including spawning sub-processes, writing to the filesystem, and coordinating agents that may have network access.
Sanitization: A multi-stage review loop involving mechanical self-checks, automated verification, and a dedicated QA cross-review agent is defined in SKILL.md to identify and remediate problematic sub-agent output before acceptance.

oma-orchestration