The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates a multi-agent system that ingests external data from company documents and messaging platforms such as Slack, Discord, Telegram, and Gmail.
Ingestion points: The application processes user messages, uploaded documents stored in packages/core/company/, and incoming messages from integrated communication channels.
Boundary markers: Instructions specify that retrieval context should be injected into the user turn rather than the cached system prompt to mitigate prompt leakage and caching issues, though strict input sanitization is not detailed.
Capability inventory: The skill allows for shell execution using Bash, file system operations via Read, Write, and Edit tools, and network communication through various integration handlers.
Sanitization: The skill documents an outbound anti-spam guard intended to dedup and rate-limit messages; however, it is noted that this guard "fails open" on internal errors and is not a strict approval gate, creating a surface for potential injection to propagate through the multi-agent pipeline.

openexecutive