testing-agentforce

Automated testing for Agentforce agents that spans quick smoke tests, persistent batch suites, and an iterative fix loop. The skill drives the Salesforce CLI directly through `sf agent preview` and `sf agent test` commands rather than a standalone Python script, and it bridges the gap between initial development and production deployment by deriving test utterances, running them, analyzing traces, and diagnosing failures.

Two testing modes are supported alongside direct action execution. Mode A (Ad-Hoc Preview Testing) runs quick smoke tests during authoring using `sf agent preview` with the `--authoring-bundle` flag, which compiles from the local `.agent` file and emits local trace files; it is best for iterative development and validating fixes. Mode B (Testing Center Batch Testing) deploys persistent test suites to the org via `sf agent test create` and `sf agent test run`, suited to regression suites, CI/CD, and team sharing. Action Execution invokes a Flow or Apex action directly through the REST `/services/data` actions endpoint for isolated debugging. When no utterances file is provided, test cases are auto-derived from the `.agent` file: subagent-based utterances, action-based utterances, a guardrail off-topic test, multi-turn scenarios, and safety probes that are always included. The plan is always presented for review before any tests run.

Traces are written under `.sfdx/agents/{BundleName}/sessions/{sessionId}/traces/{planId}.json` and analyzed with jq queries for topic routing, action invocation, grounding, safety score, enabled tools, response text, and variable updates. After safety probes the skill produces an explicit SAFE, UNSAFE, or NEEDS_REVIEW verdict. A fix loop of up to three iterations maps each failure type (such as TOPIC_NOT_MATCHED, ACTION_NOT_INVOKED, WRONG_ACTION, UNGROUNDED, or LOW_SAFETY) to a specific fix location and strategy in the agent definition.