The Agent Skills Directory

[PROMPT_INJECTION]: The skill mandates that the agent operate "SILENTLY in the background" and "don't announce it". It explicitly instructs the agent to modify user prompts, remove terms, and switch models without informing the user unless directly questioned, which significantly reduces user oversight.
[COMMAND_EXECUTION]: The skill uses shell commands via the bash tool to execute python3 higgsfield_memory.py for querying memory databases and logging results.
[PROMPT_INJECTION]: The skill utilizes relative path traversal to access and update a generation ledger and execution scripts located at ../../db/ledger/ and ../../higgsfield_memory.py. This behavior interacts with files outside of the skill's specific directory, posing a risk to file system isolation and data integrity.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting data from db/filter-memory.json and db/quality-memory.json and applying it to prompt generation. There are no boundary markers or sanitization steps provided to prevent malicious instructions within the database from hijacking agent behavior. Ingestion points: db/filter-memory.json, db/quality-memory.json. Boundary markers: Absent. Capability inventory: bash tool execution and file system writes. Sanitization: Absent.

higgsfield-recall