The Agent Skills Directory

[COMMAND_EXECUTION]: The abductive_gemini bash function in SKILL.md executes shell commands including gcloud, curl, and jq to manage authentication and interact with the Gemini API.
[EXTERNAL_DOWNLOADS]: The skill initiates network requests to googleapis.com, which is a well-known and trusted service provider, for the purpose of natural language processing.
[PROMPT_INJECTION]: The skill ingests untrusted observation data into an LLM prompt within SKILL.md. While the data is properly escaped using jq to maintain JSON integrity, the design lacks explicit delimiters or instructions to the model to disregard embedded commands, presenting a low-risk surface for indirect prompt injection. This risk is further mitigated as the resulting output is strictly parsed for numeric fields.

abductive-oracle