The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes CRM account records which could contain untrusted data. This data is used to populate human review requests in Slack, representing a managed attack surface for indirect prompt injection.
Ingestion points: CRM account properties are ingested via findRecords in infra/plays/deduplicate-accounts.ts.
Boundary markers: The skill requires manual human intervention (Slack review) for all non-exact matches, providing a significant barrier against automated exploitation.
Capability inventory: Write access is limited to the mergeRecords action on the CRM model, and network access is restricted to official CRM and Slack connectors.
Sanitization: String normalization and domain validation are performed in infra/scripts/accounts.ts to ensure consistency and strip generic values.
[EXTERNAL_DOWNLOADS]: The skill utilizes official CLI tools and packages from the vendor, specifically @cargo-ai/cli and the cargo-cdk authoring skill. These are trusted resources within the vendor's ecosystem.
[COMMAND_EXECUTION]: Local execution of development tools like cargo-ai cdk plan and validation scripts like evals/contract.mjs is required for infrastructure management and is performed within the operator's controlled environment.
[DYNAMIC_EXECUTION]: The skill uses defineScript to bundle local TypeScript logic into executable nodes and employs a contract validation script (evals/contract.mjs) that dynamically loads infrastructure resources for testing purposes. This behavior is restricted to local code within the skill package.

crm-deduplication