The Agent Skills Directory

[SAFE]: The skill provides architectural guidance for handling streaming model output. It includes implementation patterns for partial JSON buffering and stream cancellation to prevent resource exhaustion and ensures that tool execution is handled correctly.
[PROMPT_INJECTION]: The skill addresses the processing of untrusted model output streams. It provides a structured ingestion point via the consume_stream function and explicitly uses boundary markers such as the content_block_stop event to ensure tools are only dispatched with complete data. The capability to dispatch ToolCall events is mitigated by recommending json.loads only after the block buffer is verified as complete, which prevents the processing of malformed or partially injected instructions.

streaming-io