The Agent Skills Directory

[DATA_EXFILTRATION]: The skill provides instructions and code snippets to access the local configuration file ~/.zcode/v2/config.json to retrieve provider options and API keys, such as ANTHROPIC_API_KEY. This is documented as a functional requirement for the tool to interact with AI models.- [PERSISTENCE]: The documentation recommends that users add a custom shell function to their ~/.zshrc profile. This function automates the extraction and loading of credentials from the local ZCode configuration whenever the zcode command is executed.- [DYNAMIC_EXECUTION]: The skill utilizes eval in conjunction with python3 to dynamically parse JSON configuration data and export it as environment variables within the user's shell session. It also describes running local executables like zcode-acp-bridge.- [INDIRECT_PROMPT_INJECTION]: The skill describes an attack surface where the agent processes external data through file attachments and git diffs. It proactively provides security guidance for this surface:
Ingestion points: Processes untrusted data from user-provided files and git repositories via the --attach flag and tools like zcode_review and zcode_pr_review.
Boundary markers: Recommends using structural task books and explicit review prompts to guide the agent's focus.
Capability inventory: The underlying ZCode agent possesses capabilities for file modification and command execution (Write, Edit, Bash).
Sanitization: Instructs users on how to apply a "Review Mode" that physically disables write and execution tools using the --disallowed-tools flag to prevent malicious instructions in the code being reviewed from taking control of the environment.

zcode-bridge-guide