The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references a dependency on the jev-skill Python package and the jev-decide CLI tool. It follows secure practices by instructing the agent to inform the user of missing dependencies rather than attempting silent installation.
[DATA_EXFILTRATION]: In its 'Real Jev' mode, the skill is designed to send text blocks (tool outputs) to external AI providers (OpenRouter or TypeSafe) for relevance classification. This is the intended purpose of the skill. It mandates that API keys be handled as inherited environment variables and explicitly forbids printing them or including them in request files.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests and processes untrusted data from tool outputs (e.g., the blocks in assets/example.json).
Ingestion points: Data enters the context via the blocks field in JSON request files or direct tool output parsing.
Boundary markers: The skill uses structured JSON schemas to delimit processed content.
Capability inventory: The skill is advisory; it proposes keeping or dropping context but does not have the capability to execute code or write to the filesystem based on the ingested content.
Sanitization: No specific sanitization or filtering of block content is mentioned, but the shadow/advisory nature of the workflow mitigates the risk of automated obedience.

jev-context