The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill introduces an input surface for untrusted data from the Feishu/Lark messaging platform.
Ingestion points: src/channels/feishu.ts implements handleMessage, which ingests message content directly from the Feishu WebSocket event (im.message.receive_v1).
Boundary markers: The skill does not implement specific boundary markers or "ignore instructions" wrappers for the ingested content before it is passed to the agent loop.
Capability inventory: The core src/index.ts logic demonstrates that the agent has the capability to execute tasks via runContainerAgent, which performs operations within a container environment.
Sanitization: Input is parsed for JSON structure (expected for Feishu text messages), but the resulting text content is passed to the agent without further sanitization or filtering of prompt-like instructions.
Risk Factor: The skill explicitly configures the registration in SKILL.md with requires_trigger = 0, meaning the bot will process and respond to all messages in a chat without requiring an explicit mention or trigger, maximizing the exposure to potentially malicious user-provided input.

add-feishu