The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes agent transcripts to update its persistent world model, creating a potential surface for indirect prompt injection.\n
Ingestion points: The harvest.py script and hooks/stop.sh process JSONL transcript files to extract behavioral markers.\n
Boundary markers: The implementation includes significant protections; it ignores tool_result content to prevent injection via file content or command output, and requires markers to be placed at the start of a line.\n
Capability inventory: The skill maintains a persistent SQLite database (model.db) and injects summarized context into agent sessions.\n
Sanitization: The harvester uses role-based filtering, accepting authoritative oracle evidence (such as WM-VALIDATED) only from the user role, which prevents the agent from being influenced by untrusted data it might quote in its own responses.\n- [COMMAND_EXECUTION]: The evaluation script eval/grade.py uses subprocess.run to execute Python code generated by the agent within a testing context to verify the functional correctness of the agent's changes.\n- [PERSISTENCE]: The installation script scripts/install.sh registers lifecycle hooks in the agent's settings.json file, which is the required mechanism for enabling persistent codebase memory and turn-by-turn observation.\n- [DYNAMIC_EXECUTION]: The eval/run_eval.py script uses importlib.util for dynamic module loading to facilitate automated testing and validation of the skill's components.

world-model-ledger