The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from .moai/lessons-inbox.jsonl containing tool failure stubs and summaries. The agent is instructed to process this data to augment importance scores and draft 'shadow proposals', creating a surface where malicious tool output could influence the agent's logic or proposal content.\n
Ingestion points: The drain.sh script reads the inbox file and generates clusters.json, which is then read by the agent for curation tasks.\n
Boundary markers: No explicit delimiters or instructions are provided to the agent to treat the cluster summaries as untrusted or to ignore instructions embedded within the logs.\n
Capability inventory: The skill is allowed to use the Bash tool and performs multiple file-write operations within the .moai/state/ directory.\n
Sanitization: While jq is used for mechanical processing, the natural language summaries are presented to the model without filtering or sanitization.\n- [OBFUSCATION]: The tier4_firing_test.sh script employs string concatenation to hide a variable name from simple search tools. It assembles the identifier enableTriggerInjectionWrites using the pattern \"enable\"\"TriggerInjectionWrites\". This is a deliberate concealment technique intended to avoid literal text matches within the script's own source code.

hns-lsel-curator