Oracle
Read-only architecture and debugging consultant. High-IQ second opinion without execution authority.
Oracle is the consultant. Read-only, GPT-native, tuned for sustained reasoning across unfamiliar patterns, security concerns, and multi-system tradeoffs. Sisyphus delegates here when a decision needs a strong second opinion before code moves.
Default model
| Field | Value |
|---|---|
| Default | openai|github-copilot|opencode|vercel/gpt-5.5 (variant high) |
| Style | Principle-driven |
Runtime fallback chain
gpt-5.5 (high) → google\|github-copilot\|opencode\|vercel/gemini-3.1-pro (high) → anthropic\|github-copilot\|opencode\|vercel/claude-opus-4-7 (max) → opencode-go\|vercel/glm-5.1.
Tool restrictions
Read-only. Blocked: write, edit, task, call_omo_agent. Designed to advise, not implement.
When to invoke
Ask @oracle to review this design and propose an architecture.- Multi-system tradeoffs where one wrong call cascades.
- Security or threat-model decisions that benefit from sustained reasoning.
What it doesn't do
- Make changes — read-only by design.
- Delegate work — no
task/call_omo_agent. Sisyphus drives any follow-up. - Run on MiniMax — coherence loss on sustained reasoning; the substitution rules in Agent-Model Matching call this out explicitly.
Source Notes
Aligned with upstream docs/reference/features.md#core-agents, docs/reference/features.md#tool-restrictions, and src/shared/model-requirements.ts.