The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides development templates that involve filesystem interaction, such as fs.writeFile for saving configuration data in the config.ts.template. It also includes defensive programming guidance; for instance, the Permission Gate example in references/examples-gallery.md demonstrates how to implement logic to detect and block high-risk commands like sudo, rm -rf, and chmod 777.\n- [INDIRECT_PROMPT_INJECTION]: The skill defines an extension framework that ingests data from agent events, creating a potential surface for indirect prompt injection.\n
Ingestion points: Event payloads such as AgentStartedEvent.prompt and PromptSubmittedEvent.prompt provide paths for untrusted data to enter the agent context (defined in references/event-types.md).\n
Boundary markers: Delimiters are not explicitly enforced in the base templates, but the skill provides examples of logic-based blocking for dangerous inputs.\n
Capability inventory: The ExtensionContext and TaskContext interfaces allow extensions to run prompts (runPrompt), execute custom commands (runCustomCommand), and update settings (updateSettings) (found in references/extension-interface.md).\n
Sanitization: No default sanitization is provided in the templates, but the documentation includes a pattern for manual command filtering via string matching in the Permission Gate example.

extension-creator