The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill configures the agent to add and interact with an external software registry at https://registry.gascity.com/registry.toml and performs software package installations using the gc import install command.
[COMMAND_EXECUTION]: The skill instructs the agent to execute various CLI commands using ao and gc tools to manage the execution environment. This includes ao gc prepare, which writes to $CODEX_HOME/config.toml to seed trust levels, and tmux capture-pane to inspect runtime state.
[PROMPT_INJECTION]: The skill exposes an indirect prompt injection surface by requiring the agent to monitor and process data from external, potentially untrusted sources.
Ingestion points: Untrusted data enters the agent's context from tmux -L <socket> capture-pane -pt <session> (shell output), gc mail inbox (email messages), and the Gas City supervisor API ($API/runs/<run-id>).
Boundary markers: While the skill advises the agent not to infer completion from prose, it lacks explicit delimiters or instructions to ignore malicious commands embedded within the captured pane or mail text.
Capability inventory: The agent has capabilities to execute shell commands, interact with network APIs via curl, and perform file-write operations to configuration paths.
Sanitization: There is no evidence of sanitization or validation of the data ingested from shell panes or emails before it is processed by the agent.

using-gc