The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/tip_gui_bridge.py dynamically generates and executes Python code snippets via subprocess.run to interact with a local GUI environment. It also uses importlib.util to dynamically load modules from computed local paths.
[COMMAND_EXECUTION]: Both scripts/tip_gui_bridge.py and scripts/tip_gui_client.py utilize subprocess.run to execute shell commands and launch sub-processes.
[CREDENTIALS_UNSAFE]: The skill accesses a sensitive configuration file at ~/Library/Application Support/Tip/settings.json, which stores API keys and authorization headers for LLM providers. These credentials are extracted and used to make external network requests.
[INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to instructions embedded in desktop images.
Ingestion points: Desktop screenshots captured in scripts/tip_gui_bridge.py (via cmd_describe) and processed by external vision models.
Boundary markers: Absent; no delimiters or warnings are used to prevent the agent from obeying instructions found within the images.
Capability inventory: scripts/tip_gui_bridge.py has the capability to perform GUI actions (click, type, press) and write artifacts to the filesystem.
Sanitization: Absent; the descriptive output from the model is returned directly to the agent context without validation or filtering.
[DATA_EXFILTRATION]: The skill captures desktop screenshots and stores them in a local temporary workspace. It also transmits captured image data and user-associated API keys to external API endpoints as part of its vision-probing functionality.

tip-gui-skill