The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches the axe-core accessibility testing library from Cloudflare's CDN (cdnjs.cloudflare.com) during the capture phase. This is a well-known, trusted service used here for legitimate automated testing.
[COMMAND_EXECUTION]: The reporting script (scripts/build-report.py) utilizes subprocess.run to call the system's sips utility for image resizing. This execution is scoped to processing local screenshots generated during the benchmarking process.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze instruction files from other agents (contenders). While this exposes an attack surface where a malicious contender could attempt to influence the auditing agent, the skill defines a structured workflow involving isolated runs and multiple independent judge models to mitigate this risk.
Ingestion points: Reads all instruction files and references of contenders as specified in SKILL.md (Step 1).
Boundary markers: The process involves splitting readings into shards and recording line-ranged ledgers to ensure complete and discrete analysis.
Capability inventory: File reading, screenshot capture via agent-browser, and HTML generation with local subprocess calls for image manipulation.
Sanitization: The report generation script (scripts/build-report.py) employs HTML escaping and controlled markdown parsing to prevent injection into the final report output.

showdown