The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill represents a vulnerability surface for indirect prompt injection because it is designed to ingest and process data from adversarial test scenarios, including web application console logs and external service outputs.
Ingestion points: The agent reads data from browser_vision, browser_console, and outputs from external services or local installations during QA probes (documented in SKILL.md and references/manual-test-guide.md).
Boundary markers: The skill explicitly instructs the agent to 'wait for the user's explicit go-ahead' and 'user's approval' before executing destructive actions, submitting data on live targets, or expanding scope (documented in SKILL.md).
Capability inventory: The skill can modify system states (files, processes, ports, accounts) and delegate code mutations to other executors (documented in SKILL.md and references/board-fanin.md).
Sanitization: The security model relies on human-in-the-loop checkpoints and explicit approval gates for actions that could materially change external state.
[COMMAND_EXECUTION]: The skill utilizes vendor-specific CLI tools, such as omh runtime record and omh_agent_board commands, to manage task state and evidence across different workspace environments. These commands are integrated into the framework's operational workflow.

ulw-qa