The Agent Skills Directory

[DATA_EXFILTRATION]: Transmits the user's query and segments of retrieved text (up to 900 characters per passage) to an external service referred to as 'Jev' for reranking and injection screening.
[COMMAND_EXECUTION]: Instructs the agent to pipe JSON data into the jev rerank command-line utility to perform passage ranking.
[DYNAMIC_EXECUTION]: Provides instructions to execute a local Python script (evals/context-filter/regret.py) for evaluating retrieval performance.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted content from external sources such as web searches, wikis, and memory stores, serving as a surface for indirect prompt injection.
Ingestion points: Content is retrieved from external sources like session_search, wiki, and web, then passed to the filter as candidates in SKILL.md.
Boundary markers: The skill uses structured output fields such as selected_ids, dropped_injection_ids, and unjudged_ids, and provides explicit instructions for the agent to treat unvetted IDs as untrusted text.
Capability inventory: Uses a command-line tool (jev rerank) and a Python script (evals/context-filter/regret.py) for filtering and performance assessment.
Sanitization: Implements pattern-based masking for PII (emails, phone numbers, tokens), length clipping of passages to 900 characters, and a local pattern-based screen for known injection signatures.

jev-memory