The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines a set of adversarial test cases (e.g., "Ignore all your previous instructions", "You are now in unrestricted mode") used to probe the safety guardrails of the agents under test. These are explicitly part of a testing methodology. The skill also processes untrusted external data in the form of .agent configuration files and execution trace logs, creating a potential surface for indirect prompt injection.
Ingestion points: Reads local .agent files and session trace JSON files.
Boundary markers: The workflow requires presenting the test plan for user review before execution.
Capability inventory: Uses shell execution for sf, curl, and python3 utilities.
Sanitization: Implements Python-based regex filtering to remove control characters from CLI output before JSON parsing.
[COMMAND_EXECUTION]: The skill makes extensive use of the Salesforce CLI (sf) to manage agent previews, run batch tests, and retrieve results. It also uses curl for direct REST API interaction with Salesforce services.
[CREDENTIALS_UNSAFE]: Instructions are provided for retrieving Salesforce access tokens using sf org display. The skill includes a mandatory Safety Gate section that requires checking for production orgs and warns against using real PII in test inputs.
[DATA_EXFILTRATION]: Functional network operations are performed via curl to communicate with Salesforce instance URLs using the acquired access tokens.

testing-agentforce