Skip to main content

Implement “discovery as a capability”

Stub

This How-to is a stub. JARVIS has a Selection Service and uses CandidateSet artifacts, but the “discovery node” authoring surface (treat candidate-set generation as a reusable NodeType) is not yet stabilized.

Goal​

You will implement “discovery as a capability”: a reusable workflow that generates a bounded CandidateSet for a subtask and records it as a durable artifact.

When to use this​

  • You want a reusable “find viable tools” capability you can call from composites.
  • You want candidate-set generation to be auditable and replayable (durable artifacts).
Standard vs. implementation

Prerequisites​

  • A running Selection service
  • A Node Registry with inventory seeded (at least one NodeType)
  • A decision on where the discovery logic lives:
    • as a composite “discovery node” executed by Composite Executor, or
    • as a dedicated “discovery service” (not planned for v0.x).

Steps​

  1. Define a “discovery node” input schema (subtask goal + optional context/constraints).
  2. Implement discovery logic that calls Selection with:
    • subtask spec,
    • the current constraint envelope (depth/branch/budget),
    • the current policy posture (candidate filtering).
  3. Emit a durable CandidateSet artifact (and/or candidate_set_generated event).
  4. Return a reference to the candidate set (candidate_set_id) to the caller.

Verify​

  • The discovery node produces a durable CandidateSet artifact that contains:
    • candidate_set_id
    • subtask_id
    • candidates[] (each candidate contains at least a node_type_ref)

Troubleshooting​

  • Empty candidate set → inventory/policy/constraints filtered everything → validate Node Registry seeding and policy profile.
  • 5xx errors mentioning LLM → LLM provider misconfigured → verify ARP_LLM_API_KEY + ARP_LLM_CHAT_MODEL (default openai profile). For offline testing, opt into dev-mock with ARP_LLM_PROFILE=dev-mock + fixtures.
  • Candidate sets too large → enforce strict bounds at Selection and fail closed on violations.

Cleanup / Rollback​

  • None.

Next steps​