The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several local system commands through spawn, spawnSync, and execFileSync. These include vault (for secure token storage), git (for repository identity resolution), osascript (to securely prompt the user for credentials), oathtool (for generating TOTP codes), and lsof (for process cleanup by specific PID).
[INDIRECT_PROMPT_INJECTION]: The skill processes external data that could theoretically contain malicious instructions, though it is mitigated by structural validation.
Ingestion points: Ingests JSON files from the specs/ directory and scrapes text/HTML content from GitHub settings pages during browser automation.
Boundary markers: Uses a formal JSON Schema to validate input specifications and employs specific CSS selectors and ARIA roles to limit the scope of web interactions.
Capability inventory: Filesystem access for profile management and temporary storage, network communication via Chrome, and shell command execution for secret management.
Sanitization: Input specifications undergo structural validation; scraped web content is filtered using regular expressions and strict string matching to identify permissions and token IDs.
[DYNAMIC_EXECUTION]: The engine utilizes page.evaluate() to inject and run JavaScript within the browser context to interact with the GitHub UI. It dynamically constructs regular expressions and locators based on values provided in the JSON specifications (such as account names or resource owners) to drive form submission and verification logic.

gh-fine-grained-pat