/build-agentforce-action — Scaffold a complete Agentforce action

Scaffold a complete Agentforce action with a single command.

Wraps agents/agentforce-builder/AGENT.md. Produces Apex @InvocableMethod + topic YAML + agent definition + test class + starter eval.

Step 1 — Collect inputs (ask all five upfront)

Ask:

1. Action name (user-facing label)?
   Example: "Summarize Account Cases"

2. Primary sObject the action grounds on?
   Example: Account

3. Actor invoking the action?
   Example: Service Agent / Sales Rep / Customer

4. Intent (what should the action do)?
   One or two sentences. What data does it retrieve / what does it change?

5. Trust constraints?
   Pick any that apply: no-pii-in-prompt, mask-email, no-external-callout,
   require-user-confirmation, audit-every-invocation, rate-limit-per-actor.
   Add free-form constraints as needed.

If any of the five is missing, STOP and ask.

Step 2 — Load the agent

Read agents/agentforce-builder/AGENT.md fully + the Agentforce skills + evals/framework.md + the templates under templates/agentforce/.

Step 3 — Execute

Follow the 6-step plan:

  1. Classify the action (read-only / write / composite / callout)
  2. Generate the Apex action class (subclass AgentActionSkeleton, CRUD/FLS via SecurityUtils, logging via ApplicationLogger)
  3. Generate the topic YAML (classifier prompt, scope boundary, grounding sources, confirmation flag)
  4. Generate the agent definition JSON
  5. Generate the test class (including bulk + runAs + wrong-actor tests)
  6. Generate the starter golden eval

Step 4 — Deliver

  • Action summary
  • Generated files, each as a fenced block labelled with its target path:
    • Apex action class + meta
    • Test class + meta
    • Topic YAML
    • Agent meta XML
    • Golden eval markdown
  • Trust checklist (each constraint → where it's enforced)
  • Citations

Step 5 — Recommend follow-ups

  • /gen-tests if additional coverage beyond the scaffold is needed
  • /scan-security on the action Apex before promoting to production
  • Deploy the eval file under evals/golden/ and run evals/scripts/run_evals.py

What this command does NOT do

  • Does not deploy.
  • Does not run the eval.
  • Does not modify existing agents / topics — creates new ones.
  • Does not allow callouts when no-external-callout is in the trust constraints.