The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates the processing of task objects which include messages and state history from external client agents, creating an indirect prompt injection surface.
Ingestion points: Untrusted data enters the agent context through the 'messages' and 'metadata' fields of the Task object, as well as via the message/send and message/stream endpoints described in SKILL.md.
Boundary markers: The instructions lack any requirement for delimiters or specific system messages to help the agent distinguish between task data and its own core instructions.
Capability inventory: The skill allows the use of powerful tools including Bash, WebFetch, WebSearch, and file system access (Read, Write, Edit), increasing the potential impact of a successful injection.
Sanitization: There are no procedures defined for validating or sanitizing the content of task messages before they are processed by the agent.
[EXTERNAL_DOWNLOADS]: The skill directs the agent to fetch documentation and search for implementation examples from external sources to guide its work.
Evidence: Fetches the specification from https://a2a-protocol.org/latest/specification/ and searches github.com/a2aproject for task handling samples.
Context: These operations are intended for documentation retrieval and do not involve the direct execution of remote code or scripts.

a2a-task-lifecycle