The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests performance metrics and script metadata from a browser environment, creating a potential surface for indirect prompt injection if the profiled page contains adversarial content.
Ingestion points: Browser PerformanceObserver entries (long animation frames, events, layout shifts) processed in observe.js.
Boundary markers: The instructions in SKILL.md define a specific reporting format (`1. file:line
[rule-id]...`) to structure the agent's output.
Capability inventory: Uses browser evaluation capabilities to run performance monitoring scripts.
Sanitization: The describe function in observe.js limits captured element metadata to tag names and simplified CSS classes/IDs to reduce data leakage.
[DATA_EXFILTRATION]: The observe.js script collects script metadata, including source URLs and function names, from the environment it profiles. This could expose internal application structure or sensitive tokens if they are embedded in script URLs.
Evidence: scripts: entry.scripts.map((script) => ({ source: script.sourceURL, function: script.sourceFunctionName, ... })) in observe.js.

frames