Skip to main content

Implement the mapper

Stub

This How-to is a stub. JARVIS has a working mapper (Selection → binding → arg-gen), but the user-facing authoring story is still evolving.

Goal​

You will map each subtask to a bounded CandidateSet, bind to a specific NodeTypeRef, and generate the concrete input for the chosen node.

When to use this​

  • You want a clean separation of responsibilities:
    • Selection produces a bounded menu (CandidateSet),
    • the composite executor performs binding + argument generation.

Prerequisites​

  • A running Selection service that can query Node Registry inventory
  • A binding strategy (deterministic first; LLM-assisted later)
  • An argument generation (arg-gen) schema for each target NodeType input

Steps​

  1. For each subtask, call Selection to get a bounded CandidateSet.
  2. Choose one candidate (binding decision) and record it as an artifact/event.
  3. Run arg-gen to produce the concrete input JSON for the chosen NodeTypeRef.
  4. Create the child NodeRun via Run Coordinator.

Verify​

  • Each subtask produces:
    • a CandidateSet artifact,
    • a binding decision artifact,
    • an arg-gen artifact (inputs for the chosen node).
  • Child NodeRuns are created with schema-valid inputs.

Troubleshooting​

  • Empty candidate set → inventory/policy/constraints filtered everything → validate Node Registry seeding + Selection configuration.
  • Binding picks a wrong node → keep binding deterministic and conservative until eval/recovery loops are mature.
  • Arg-gen produces invalid inputs → fail closed; trigger recovery (remap or retry with bounds).

Cleanup / Rollback​

  • None.

Next steps​