The Agent Skills Directory

[DATA_EXFILTRATION]: The skill's primary function is to send the content of local files and logs to external worker models (such as DeepSeek, Gemini, Qwen, or GLM) for summarization. This constitutes a deliberate data flow of local file content to external API endpoints.
[INDIRECT_PROMPT_INJECTION]: The skill implements a pattern where a worker model processes untrusted external data (files and logs) and returns a summary to the primary agent. Malicious instructions embedded in the processed files could potentially influence the agent's behavior through the generated summaries.
Ingestion points: Untrusted files and logs read via the bulk-read command (e.g., source code, configuration files, or command output).
Boundary markers: The skill does not mention the use of delimiters or specific instructions to the agent to ignore embedded commands within the shunted content.
Capability inventory: The tool calls involve reading file contents and invoking external CLI tools for model inference.
Sanitization: There is no evidence of sanitization or filtering of the input files or the resulting worker model summaries.
[COMMAND_EXECUTION]: The skill references and encourages the execution of external CLI tools and worker commands (e.g., bulk-read, deepseek --flash, gemini-api --flash-lite, qwen3, glm) to perform its summarization tasks.

context-shunt