The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill contains an UPDATE FIRST mechanism in SKILL.md that instructs the agent to check a remote manifest and automatically replace the local skill directory with content downloaded from a GitHub zip archive.
[COMMAND_EXECUTION]: The instructions for end-to-end testing in T2-test-runners.md specify that the agent should install browser automation software and browsers without asking for user confirmation, which bypasses oversight on environment modifications.
[DYNAMIC_EXECUTION]: The agent is directed to generate and execute a reporting script and corresponding shell or batch wrappers within the target repository to process test results and build HTML reports.
[EXTERNAL_DOWNLOADS]: The skill makes several requests to external repositories, including checking manifests on raw.githubusercontent.com and downloading release assets from GitHub APIs.
[INDIRECT_PROMPT_INJECTION]: The skill processes clinical and administrative data received from external NHCX and ABDM endpoints. Fields such as disposition and remarks are ingested and represent a surface for instruction injection from external parties. Ingestion points: Gateway inbound route and Callback door. Boundary markers: Relies on JSON structure of FHIR bundles; no specific delimiters for natural language fields are defined. Capability inventory: Performs network requests, writes to local files, and executes shell commands for testing. Sanitization: Structural validation is performed, but content within descriptive fields is not sanitized for potential prompt instructions.

nhcx-communication