The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to execute a Python script (install_hermes_qq.py) that performs invasive patching of the host repository's source code, altering the fundamental logic of the application.
[DYNAMIC_EXECUTION]: The verification documentation (references/verification.md) instructs the user/agent to execute a Python code block directly via the shell using a heredoc (`python
<<'PY'`), which is a technique for running unverified code at runtime.
[PERSISTENCE]: By modifying core application files such as gateway/run.py and gateway/config.py, the skill ensures that its changes and the new QQ platform support persist across agent sessions and updates.
[INDIRECT_PROMPT_INJECTION]: The skill introduces a new ingestion point for untrusted data via the QQ platform adapter.
Ingestion points: assets/qq.py receives messages from the QQ WebSocket API.
Boundary markers: None identified in the provided adapter code to distinguish user content from instructions.
Capability inventory: The adapter has the ability to send messages and upload various file types (.json, .zip, .pdf) back to the platform.
Sanitization: The adapter performs basic parsing but does not include safety filtering for the message content before passing it to the agent gateway.
[DATA_EXFILTRATION]: The script scripts/install_hermes_qq.py patches gateway/platforms/base.py to include sensitive file extensions like .json, .csv, and .zip in the agent's media extraction regex. This expands the data exfiltration surface, allowing the agent to identify and potentially send sensitive data files to external chat platforms.

hermes-qq