The Agent Skills Directory

[SAFE]: The skill is an instructional architectural guide describing a design pattern for multi-agent orchestration. It does not contain executable scripts, obfuscated code, or network-enabled logic.
[DATA_EXPOSURE]: The skill does not access sensitive local files or hardcode any credentials. It focuses on development workflow tasks such as fixing test files.
[COMMAND_EXECUTION]: The skill mentions a platform-native Task() function used to spawn sub-agents. This is treated as intended architectural documentation rather than arbitrary command execution.
[PROMPT_INJECTION]: No instructions were found that attempt to bypass AI safety guardrails, override system instructions, or extract system prompts.
[INDIRECT_PROMPT_INJECTION]: The pattern involves ingesting external data (such as test failures and error logs) and passing it to sub-agents. This establishes an attack surface for indirect prompt injection if the logs contain untrusted external content.
Ingestion points: Test logs, error messages, and file contents are used to populate sub-agent prompts.
Boundary markers: The examples show basic formatting but no explicit security delimiters for untrusted input.
Capability inventory: Sub-agents spawned via Task() possess file-read and file-write capabilities.
Sanitization: No explicit sanitization or filtering of error logs is mentioned in the pattern guide.

dispatching-parallel-agents