Prometheus
Strategic planner. Interviews you, surfaces ambiguity, and writes a plan before execution touches the codebase.
Prometheus is the planning lane. Press Tab in the OpenCode UI or type @plan "<your task>" from Sisyphus to enter Prometheus mode. It interviews you, researches your codebase to ground the questions, surfaces edge cases you hadn't considered, and writes a .sisyphus/plans/*.md plan. Then /start-work activates Atlas to execute it.
Default model
| Field | Value |
|---|---|
| Default | anthropic|github-copilot|opencode|vercel/claude-opus-4-7 (variant max) |
| Style | Dual-prompt — Claude (mechanics-driven, ~1,100 lines) or GPT (principle-driven, ~121 lines) |
| Switch | isGptModel() at runtime |
When the resolved model is GPT-5.4 or GPT-5.5, Prometheus loads the GPT prompt. Otherwise it loads the Claude prompt. You don't choose; the runtime does.
Runtime fallback chain
claude-opus-4-7 (max) → openai\|github-copilot\|opencode\|vercel/gpt-5.5 (high) → opencode-go\|vercel/glm-5.1 → google\|github-copilot\|opencode\|vercel/gemini-3.1-pro.
When to invoke
- Multi-day projects with unclear scope.
- Critical production changes that need a documented decision trail.
- Complex refactors where "what to build" is half the work.
Skip Prometheus when the task fits in one prompt or one file — ulw is faster.
Tool surface
Read + research tools plus the task tool for delegating spike research to specialists. Cannot edit production code directly; that's by design.
Output constraint enforced by the prometheus-md-only hook — plans must be markdown.
What it doesn't do
- Edit production code — its lane is plans, not commits. Atlas executes the plan.
- Stop after one question — it iterates until scope is "Decision Complete."
Source Notes
Aligned with upstream docs/reference/features.md#planning-agents and src/shared/model-requirements.ts. Why Claude and GPT need different prompts is covered in Agent-Model Matching.