Skip to main content

Implement the decomposer

Stub

This How-to is a stub. JARVIS implements decomposition in Composite Executor, but we have not yet stabilized a public “decomposer authoring” surface beyond the internal planner prompt + JSON schema.

Goal​

You will implement bounded decomposition: turning a composite goal into a small set of subtasks while enforcing depth/branch limits.

When to use this​

  • You are authoring a planner-style composite node (general planner, coder planner, domain planner).
  • You want deterministic safety bounds (no unbounded action space).

Prerequisites​

  • A running Composite Executor implementation
  • A constraints model (examples: max depth, max subtasks, max decomposition rounds)
  • A structured planner output schema (strict JSON)
  • An LLM provider configured (default openai, or opt into dev-mock for offline tests)

Steps​

  1. Define the planner output schema as strict JSON (subtasks list with stable IDs).
  2. Write a planner system prompt that:
    • repeats the hard bounds explicitly,
    • disallows side quests,
    • emits only schema-valid JSON.
  3. Enforce bounds after parsing:
    • reject outputs that violate bounds,
    • record the failure as an artifact/event,
    • (optional) retry with a bounded “repair” prompt.

Verify​

  • The decomposer output is schema-valid JSON.
  • Decomposition never exceeds the configured max subtasks/depth/rounds.

Troubleshooting​

  • Invalid JSON → tighten schema + add bounded repair retries.
  • Too many subtasks → enforce bounds after parsing and fail deterministically.
  • Planner ignores constraints → treat violations as errors; do not “best effort” execute.

Cleanup / Rollback​

  • None.

Next steps​