adlc-self-orchestrate

name	adlc-self-orchestrate
description	How an Antigravity session decomposes a large build-out into a ticket DAG and drives the agb fleet itself (recursive orchestration). Use when asked to "build X in parallel", "run a fleet", "orchestrate this with agb", or when a task is too large for one session.

Self-Orchestration with agb

You can multiply yourself: decompose the work into a ticket DAG, then run agb so parallel agy workers build it while you supervise. Control flow belongs to agb (deterministic); you supply the decomposition and the escalation judgment.

When

Use this only for substantial work (3+ independently-buildable parts). Single bounded tasks: just do them.

Procedure

0. Prefer compiling an existing plan. Planning belongs to Antigravity's plan phase, not to you ad hoc: if this work was planned in a planning conversation (this session or the desktop app), a brain artifact (implementation_plan.md) already exists. Compile it instead of decomposing by hand — npx agb brains to find it, then npx agb plan <id> <repo>, which converts, gates (overlap, coldstart, parallax, premortem), and writes a provenance-stamped plan.json. If gates report blocking findings, refine the plan (the markdown, not the JSON) and recompile. Steps 1–3 below are the fallback for when no brain artifact exists.
1. Foundation first. Identify shared surface (schemas, types, shared utilities). Build and commit it to main BEFORE the fan-out — parallel workers consume the foundation, never invent it. Foundation paths go into every ticket's rails (read-only).
2. Decompose into tickets (adlc schema). Each ticket must be executable by a fresh agent from its body alone — full file paths, exact acceptance criteria, named gate commands. Partition scopes so no two tickets share files; shared needs = a foundation item or an edge.
3. Write the plan file plan.json:

{
  "repo": "/abs/path/to/repo",
  "base": "main",
  "gate": { "build": "npm run typecheck", "test": "npm test" },
  "tickets": [
    { "id": "T1", "title": "...", "body": "self-contained spec",
      "scope": ["src/feature-a/**"], "rails": ["test/contracts/**"],
      "edges": [{ "to": "T3" }], "tier": "cheap|mid|frontier",
      "pool_hint": "auto" }
  ]
}

Tier by escape cost, not prestige: cheap (Gemini Flash) where gates catch everything; mid (Claude Sonnet / Gemini Pro) default; frontier only for contracts/migrations with thin rails.

Pool hint routing: pool_hint ∈ gemini | claude | auto. Pair builder pool hints so prosecution runs cross-family (e.g. gemini builder + claude prosecutor) for maximum refutation power. 4. Validate, then run:

npx agb validate plan.json
npx agb run plan.json        # exit 0 all merged, 2 some failed
npx agb status /abs/path     # live dashboard from another terminal

5. Supervise, don't babysit. agb already does two-strike regeneration, cross-model prosecution, and gate-blocked merges. Your job is only the escalations: tickets that fail twice mean the ticket is wrong — rewrite the ticket (smaller scope, more explicit body), don't coach the agent.
6. Report from .booster/report.json: merged, failed (with reasons), requests per quota pool.

Quota safety

• Width is capped per pool by agb; do not raise caps above the calibrated ceilings without re-probing (npx agb probe).
• A run of N tickets costs roughly 2N–4N requests (build + prosecution + retries). Check that against your weekly budget before launching.