The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages workflows driven by LLM outputs, which constitutes a potential injection surface. However, the skill provides and demonstrates robust mitigation patterns:
Ingestion points: Processes model-generated retrieval tasks and action plans in examples/instant_retrieval_overlap.py and examples/response_fanout.py.
Boundary markers: Utilizes Agently's structured output controls to enforce schema compliance on model responses.
Capability inventory: Actions are limited to workflow routing and state updates within the execution context; no arbitrary shell execution or filesystem access is exposed to the model.
Sanitization: Implements strict developer-authored validation functions (e.g., validate_plan) to verify the integrity and constraints of ingested data before processing.
[DYNAMIC_EXECUTION]: The skill includes mechanisms for execution state persistence (save and load). As documented in references/recovery.md, these features are designed for safety by excluding live coroutines, secrets, and system-level handles from serialization, requiring explicit host-side resource re-injection to prevent the execution of untrusted or stale state handles.

agently-triggerflow