The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Fetches and executes installation scripts from well-known services (bun.sh) and vendor-owned resources (plannotator.ai). These downloads are used to establish the local runtime environment for planning and execution.
[COMMAND_EXECUTION]: Extensively uses embedded Python snippets and shell scripts to perform complex orchestration tasks. This includes managing git worktrees, tracking execution state in .omc/state/omg-state.json, and configuring platform-specific hooks in user directory settings (e.g., ~/.claude/settings.json, ~/.codex/config.toml).
[DATA_EXFILTRATION]: Performs local network activities confined to the loopback interface (127.0.0.1). These operations include port probing via /dev/tcp to verify tool availability and polling a local agentation MCP server at localhost:4747 for visual UI annotations.
[PROMPT_INJECTION]: The skill ingests untrusted data from implementation plans (plan.md) and external UI feedback, creating a surface for indirect prompt injection.
Ingestion points: Reads from local Markdown files and the http://localhost:4747/pending API.
Capability inventory: Possesses significant capabilities across all scripts, including file writing, subprocess execution, and git manipulation.
Boundary markers: Implements mandatory human-in-the-loop approval gates (e.g., plannotator visual review) before moving from the planning phase to execution, significantly reducing the risk of automated injection success.
Sanitization: Relies on the host agent's native guardrails and user verification during the planning and verify phases.

omg