The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and analyze external rendered content including screenshots, terminal UI frames, and browser interaction traces. This ingested data could contain hidden instructions or malicious patterns intended to manipulate the agent's QA verdict.
Ingestion points: SKILL.md specifies that the agent should analyze "render/capture evidence", "TUI frames", and "browser screenshots".
Boundary markers: The skill requires an "exact capture-to-target lineage" and uses a "visual-verdict-contract.md" to structure outputs.
Capability inventory: The agent generates verdicts (PASS, REVISE, BLOCK) and records runtime events using the omh command-line tool.
Sanitization: The skill explicitly instructs to "redact credentials/PII from captures", though it lacks specific mechanisms to sanitize instructions embedded within the visual content itself.
[COMMAND_EXECUTION]: The skill utilizes an internal execution harness to track its runtime state.
Evidence: SKILL.md contains the shell command omh runtime record --skill visual-qa --harness visual-qa --status started used for internal telemetry and state management.

omh-visual-qa