The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill coordinates workflows that ingest untrusted data from external sources (web research, paper reviews). This creates a vulnerability surface for indirect injection.
Ingestion points: Data enters the context from external research skills such as omh-web-research and ulw-research.
Boundary markers: The skill explicitly defines a 'not-evidence boundary' and uses a source_inbox/v1 structure with separate buckets (raw, processed, brief, conflict, verification) to maintain isolation between data processing stages.
Capability inventory: The workflow includes the ability to manage task persistence via cronjob_manage and read local configuration files via skill_view.
Sanitization: It requires 'observed evidence' from tool results before any state transition and utilizes verification buckets to flag conflicts, reducing the likelihood of the agent obeying instructions embedded in retrieved content.
[SAFE]: Scheduling operations via the cronjob_manage tool are restricted to authenticated tool outputs and require explicit approval, preventing unauthorized persistence.
[SAFE]: The shell commands presented in the skill are documentation examples for the omh CLI and do not trigger automatic or hidden execution.

omh-research-department