The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data, which is an inherent attack surface for indirect prompt injection. However, the skill implements strong mitigation strategies.
Ingestion points: External data enters the system through scripts/import_metrics.mjs, which reads JSON payload files (e.g., payload.json) prepared from casebase metadata.
Boundary markers: The skill enforces a mandatory human-in-the-loop review queue. Every management insight must be approved, revised, or blocked by a human partner before any further action or export occurs.
Capability inventory: The skill has the capability to write to the local file system (via scripts/export_management_report.mjs) and write to the Busabase data backend. It does not have arbitrary command execution or unrestricted network access.
Sanitization: The skill instructions explicitly mandate the use of anonymized metadata only and forbid the importation of raw confidential documents or privileged notes into the dashboard.
[REMOTE_CODE_EXECUTION]: The skill executes local automation scripts (import_metrics.mjs, execute_decisions.mjs, export_management_report.mjs) to manage the data lifecycle. These scripts are trusted local components that perform strictly defined operations within the skill's directory and do not execute remote code from untrusted sources.
[DYNAMIC_EXECUTION]: The AirApp UI utilizes dynamic imports in content/kelly-legal-firm-radar-app/app/js/providers/index.js to switch between demo and production data providers. This is a standard architectural pattern for feature toggling and does not involve executing arbitrary strings or untrusted code.

kelly-legal-firm-radar