The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary shell commands defined in the DONE-CRITERIA. While this is the intended primary functionality for verifying task completion, it allows for the execution of any command the agent is authorized to run.
[INDIRECT_PROMPT_INJECTION]: The skill accepts user-supplied objective statements, criteria, and "NOT-DONE-YET" guardrails which are stored in a state file and subsequently injected into the prompts of downstream agents during each iteration cycle.
Ingestion points: Objective statements, criteria, and guardrails are captured in Phase 1 (SKILL.md) and stored in .objective/<slug>/state.md.
Boundary markers: The skill does not explicitly define sanitization or boundary markers (like XML tags) when interpolating these fields into the /do router dispatch prompts.
Capability inventory: The skill has access to powerful tools including Bash, Write, Edit, and the ability to dispatch further agents via Task.
Sanitization: No specific sanitization logic is described for the user-supplied text before injection into the workflow.
[PERSISTENCE]: The skill utilizes ScheduleWakeup for session-scoped looping. It includes explicit logic to detect requests for permanent persistence (e.g., cron jobs) and requires a specific OWNER-APPROVED-PERSISTENCE authorization phrase before routing to persistence-related tools, mitigating unauthorized background execution.

objective-loop