The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze untrusted frontend artifacts, including DOM attributes and semantic trees, which provides a surface for indirect prompt injection. Attackers could embed malicious instructions within the web pages or components being analyzed to trick the agent into reporting a false 'PASS' or bypassing safety checks.
Ingestion points: Artifacts provided via URLs, build numbers, or local paths as described in SKILL.md and references/acceptance-matrix.md.
Boundary markers: The skill relies on 'Invariants' (e.g., 'Evidence over Narrative') and a 'Safety and Compliance Boundary' to mitigate risks, though it lacks explicit technical delimiters for external data.
Capability inventory: The agent can influence the release process by outputting PASS/FAIL/BLOCKED verdicts.
Sanitization: The skill mandates the use of synthetic test data and instructs the agent to check screenshots and logs for sensitive information before inclusion in reports.
[COMMAND_EXECUTION]: The skill includes a test suite and a shell wrapper that execute local commands to perform report validation.
Evidence: tests/test_qa_gate_contract.py uses subprocess.run() to invoke the Python interpreter for executing the report checker script during testing. Additionally, scripts/check-qa-report.sh uses exec python3 to wrap the validator logic. These are standard utility patterns but represent a local command execution surface.

frontend-qa-gate