The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates the execution of external coding CLIs through platform-native bridge tools like omh coding run and omh coding fanout dispatch. This capability is tightly controlled by instructions that forbid auto-dispatch, requiring an explicit user command and a prior successful 'readiness probe' (executor_readiness/v1) before any code is executed.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input, including task descriptions from the user and skill sets discovered from external coding profiles, to compose prompts for downstream agents. To mitigate injection risks, it enforces a rigid 10-section prompting contract, uses explicit 'evidence boundaries,' and performs sanitization by discarding raw skill descriptions in favor of verified invocation strings.
Ingestion points: Task/unit descriptions in SKILL.md, user interjections during runs, and external executor skill metadata discovered via omh coding executor-skills.
Boundary markers: Utilizes a mandatory 10-section contract (Goal, Do, Don't, etc.), a 'Docs consulted' block, and a 'claim_boundary' to delimit untrusted content.
Capability inventory: Execution of coding CLIs via omh coding run and omh coding fanout dispatch in SKILL.md; state tracking and status polling.
Sanitization: Explicitly forbids carrying external description text into composed prompts, allowing only name and invocation strings to pass through.
[DYNAMIC_EXECUTION]: The skill dynamically generates prompt strings and CLI arguments (e.g., --model, --allowedTools) based on user preferences and environment discovery. It employs a 'cache-stable' composition technique, splitting the generated output into a static head and a dynamic tail to ensure predictable execution patterns.

ulw-maestro