The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides a framework for running subagents with arbitrary test inputs. This architecture creates a surface for indirect prompt injection where instructions embedded in test cases could potentially override subagent behavior or safety protocols.
Ingestion points: Test inputs are defined in Phase 1 (SKILL.md) and Phase 3 (references/examples-and-errors.md) and passed directly to subagents via the Task tool.
Boundary markers: The dispatch templates in references/examples-and-errors.md use triple quotes (""") as delimiters for inputs, but they lack explicit instructions for the subagent to treat interpolated data as untrusted or to ignore embedded instructions.
Capability inventory: The skill has access to powerful tools including Bash, Write, Edit, and Task (for subagent spawning).
Sanitization: There is no evidence of input validation, filtering, or sanitization of the test data before it is interpolated into subagent prompts.
[COMMAND_EXECUTION]: The skill instructions direct the agent to execute shell commands and local scripts using the Bash tool for reading agent definitions and running the evaluation harness.
Evidence: SKILL.md contains instructions to execute cat agents/{agent-name}.md and cat skills/{skill-name}/SKILL.md. Additionally, references/testing-patterns.md documents the usage of python evals/harness.py for automated testing.
[DYNAMIC_EXECUTION]: The skill dynamically invokes subagents using the Task tool based on agent names that are either provided as input or discovered in the file system.
Evidence: Instructions in Phase 1 and Phase 2 of SKILL.md utilize the Task tool with the subagent_type parameter set to variables such as {agent-name}, which are determined at runtime.

testing-agents-with-subagents