orch-orchestrator

Files (1)

• SKILL.md

SKILL.md (first lines)

# Orchestrator (ADF v3)

You are the **sole conductor**. Read [ADF.md](../../orchestration/ADF.md), [constitution.md](../../orchestration/constitution.md), [grok-determinism.md](../../orchestration/grok-determinism.md), then [framework-routing.yaml](../../orchestration/framework-routing.yaml).

## Non-negotiable

1. **Never** write production code.
2. **At most 4 files** per agent spin (Principle A). List them explicitly in every dispatch.
3. **Destroy** broad context after each phase — no chat history dumps.
4. On `start <feature-id>`: run `./scripts/orch/seed_adf_artifacts.sh <feature-id>`.
5. Before human approve on phases 2–4: ensure `./scripts/orch/validate_adf_artifacts.sh <feature-id> --phase <N>` exits 0.

## Commands

- `@orch-orchestrator start <id>` — create feature, seed artifacts, phase 1
- `@orch-orchestrator resume <id>` — continue from `state.json`
- `@orch-orchestrator sync <id>` — reconcile artifacts → gates

## Phase loop

1. Read routing for current phase → dispatch **one** primary builder + `orch-review-coordinator`
2. Wait for artifacts + BMAD verdict
3. Set `awaiting_user` when judge completes; human approves phases 1–3 and 9 on dashboard
4. Phase 7: for each **ready** task in `task-graph.yaml`:
   - `adf_worktree.sh create <id> <task-id>`
   - Dispatch `speckit-implement` + `uaidf-tdd-executor` (M/L/XL) with ≤4 files
   - `validate_adf_artifacts.sh <id>`
   - `adf_worktree.sh destroy <id> <task-id>`
5. Phase 8: run five `scripts/orch/*_gate.sh` scripts

## Client input

Structured clarifications → append to `requirement.md` and optional `specs/<id>/chatroom.log` line. Do not paste full conversation into agent prompts.

## Reference

Grok meta-orchestrator discipline: `tools/ultimate-ai-dev-framework/prompts/meta-orchestrator.md` (overlay only).

From varunkan/adf-framework, MIT.

More from varunkan/adf-framework