The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides a framework for testing AI agent harnesses by processing agent-generated content and tool inputs, which creates a surface for indirect prompt injection.
Ingestion points: The runHook (tool_input parameter) and runHarnessTest (transcript parameter) functions in references/writing-tests.md ingest data that may originate from an AI agent or external source.
Boundary markers: Documentation in references/writing-tests.md highlights the use of sandboxing for untrusted hooks ({ trusted: false }) and bubblewrap confinement.
Capability inventory: The vigiles framework includes capabilities to execute shell commands and scripts via runScript, runHook, and runHarnessTest as described in SKILL.md and references/writing-tests.md.
Sanitization: The skill describes the use of a bubblewrap-based sandbox to isolate execution, blocking unauthorized file writes and network egress for untrusted code.
[DYNAMIC_EXECUTION]: The skill uses the vigiles framework to execute shell scripts and hooks for testing purposes.
Evidence: references/writing-tests.md includes examples of using runScript to execute bash scripts (e.g., bash scripts/check-links.sh).
Evidence: references/writing-tests.md demonstrates using runHook to execute hook commands derived from the project under test.
[EXTERNAL_DOWNLOADS]: The skill involves downloading dependencies and referencing external CI components.
Evidence: SKILL.md instructs the user to install the vigiles and @anthropic-ai/claude-code packages via NPM.
Evidence: SKILL.md references the zernie/vigiles@v1 GitHub Action for use in CI workflows.
[COMMAND_EXECUTION]: The skill documentation provides commands to run the testing and evaluation tools.
Evidence: SKILL.md specifies the use of npx vigiles test and npx vigiles eval to perform test runs and model evaluations.

test-harness