Hephaestus
GPT-native autonomous deep worker. Give it a goal, not a recipe.
Hephaestus runs on GPT-5.5. Give it a goal — exploration, research, end-to-end execution all happen without hand-holding. It is built for principle-driven prompts, the opposite of Sisyphus's mechanics-driven style. The name comes from the Greek god of forge and craftsmanship: the legitimate craftsman, born from necessity.
Default model
| Field | Value |
|---|---|
| Default | openai|github-copilot|venice|opencode|vercel/gpt-5.5 (variant medium) |
| Style | Principle-driven — concise XML-tagged prompt |
| Provider requirement | Requires a GPT-capable provider — single-entry chain, no Claude fallback |
Runtime fallback chain
GPT-5.5 medium only. There is intentionally no Claude or non-GPT fallback. The no-hephaestus-non-gpt hook enforces this — Hephaestus must run on GPT.
When to invoke
Switch to Hephaestus explicitly when the work needs:
- Deep architectural reasoning across many files.
- Complex debugging that benefits from goal-oriented exploration.
- Cross-domain knowledge synthesis without hand-holding.
Sisyphus's task(category: "deep") also routes deep work, but Hephaestus is the named lane for "this task is purely GPT-native deep work."
Tool surface
Full toolset. Runs autonomously, so it can spend many tokens on exploration before producing output.
What it doesn't do
- Run on Claude or MiniMax models — both are wrong fits for GPT-native autonomy. The hook blocks this.
- Replace Sisyphus's orchestrator role — Sisyphus delegates work; Hephaestus completes a single goal.
Source Notes
Aligned with upstream docs/reference/features.md#core-agents and src/shared/model-requirements.ts. Substitution policy lives in Agent-Model Matching.