| name | adg-backlog-decompose |
| description | Decompose a feature or intent into small, context-bound backlog items in the SQL backlog, for a planning model (Opus or Sonnet). Use before a swarm of workers: turn one feature into tightly scoped, logged items, each with a write scope and a context profile, so cheaper sub-agents can each work one item in isolation and total token use drops. |
ADG Backlog Decompose (planner)
You are the planner. You hold the feature view ONCE and turn it into many small, logged,
context-bound items so cheaper workers can each take one without loading the whole feature.
Loop
- Read the feature view once (this is your only wide context):
npm run context:feature -- --feature <ID> --workflow agentic-tooling --format toon,
and the lineage if the feature is elicited:
npm run elicitation:graph -- --feature <ID> --format toon.
- Split the feature into items that are each one coherent change, with a tight title, a
named write scope (the files it may touch), and a context profile (
quick-ui,
delivery-slice, or agentic-tooling). Smaller is better: a worker should need only
its own item packet (about 1.3k tokens) to act.
- Write the items into the SQL backlog seed (do not invent a parallel Markdown list),
rebuild, and confirm they are queryable: edit
data/seed/backlog.seed.json (the host
seed; the demo uses data/seed/backlog.demo.seed.json), then npm run setup (or
setup:demo) and npm run backlog:validate.
- Emit the worklist for the swarm:
npm run agent:loop -- --feature <ID> --max-items <N> (each row carries the item's
context:item command and the bounded-work constraints).
Rules
- Keep item write scopes disjoint where possible, so workers can run in parallel without
colliding.
- A sensitive item (auth, RBAC, schema, migration, guardrails, audit, billing, production)
is an L3 item; flag it so the orchestrator gives it a stronger model and the right gate.
- Keep current state derived from events; never hand-edit item status.
- Run
npm run elicitation:validate and resolve any "uncovered intent" advisory gap, so no
requirement is decomposed into work without a covering acceptance criterion.