The Agent Skills Directory

[SAFE]: The skill provides documentation and structural implementation for agent loop patterns aimed at preventing runaway behavior and ensuring task completion.
[INDIRECT_PROMPT_INJECTION]: The loop patterns (specifically Ralph Wiggum and Manage-loop) operate by analyzing agent output and external state files to make control-flow decisions.
Ingestion points: manage-loop-hook.py and the Ralph Wiggum pattern read agent transcripts from stdin; manage-continuation.json is read from the agent-memory directory.
Boundary markers: The skill utilizes scripts/schemas/manage-continuation.schema.json to validate the structure of the continuation signals.
Capability inventory: grind-loop.sh acts as a supervisor for process management and isolation; hooks utilize exit codes to signal session continuation or termination to the agent platform.
Sanitization: Validates continuation data against a fixed schema; transcript-based termination relies on specific sentinel word matching.

autonomous-loop-patterns