Atlas
Plan executor. Activates on /start-work and turns a Prometheus plan into verified tasks.
Atlas is the todo orchestrator. It reads a Prometheus plan, splits it into the right shape of tasks, delegates to specialized subagents, accumulates learnings between steps, and verifies each completion independently. The lane is execution, not planning — Prometheus already decided what to build; Atlas drives it to done.
Default model
| Field | Value |
|---|---|
| Default | anthropic|github-copilot|opencode|vercel/claude-sonnet-4-6 |
| Style | Dual-prompt — Claude or GPT, switched by isGptModel() |
Sonnet is fast and cheap enough for todo management without giving up Claude-family discipline. When GPT-5.5 takes over via fallback, Atlas auto-switches to the GPT prompt.
Runtime fallback chain
claude-sonnet-4-6 → opencode-go\|vercel/kimi-k2.6 → openai\|github-copilot\|opencode\|vercel/gpt-5.5 (medium) → opencode-go\|vercel/minimax-m2.7.
When to invoke
You don't usually invoke Atlas directly — /start-work [plan-name] does it for you on the latest Prometheus plan. Direct invocation is for resuming a previously interrupted plan.
Tool surface
Has task_* tools (task system enabled by default). Cannot use task/call_omo_agent to delegate work to other agents — its only fan-out is the task system, which keeps the execution flat and accountable.
What it doesn't do
- Generate the plan — that is Prometheus's lane.
- Re-delegate to other orchestrators — Atlas executes; subagents specialize.
- Stop when a single task fails — the
todo-continuation-enforcerhook keeps it moving.
Source Notes
Aligned with upstream docs/reference/features.md#orchestration-agents and src/shared/model-requirements.ts. The full plan-then-execute flow lives in Orchestration overview.