feat(hephaestus): support custom OpenAI-compatible providers

[imwxc]

Summary

• Remove requiresProvider restriction from Hephaestus to support custom OpenAI-compatible providers
• Allow users to configure Hephaestus with any model via oh-my-opencode.json
• When allow_non_gpt_model: true is set, a warning toast is shown but the agent is not blocked

Changes

1. Remove Provider Restriction (src/shared/model-requirements.ts)

• Removed requiresProvider: ["openai", "opencode"] from Hephaestus
• Users can now configure any OpenAI-compatible provider (OpenRouter, LiteLLM, custom proxies)
• GPT remains the default via existing fallbackChain

2. Hook Behavior (src/hooks/no-hephaestus-non-gpt/hook.ts)

• Default behavior: Block non-GPT models with error toast, switch to Sisyphus
• With allow_non_gpt_model: true: Show warning toast but allow the model to proceed

Usage Example

Users can now configure Hephaestus with custom thinking-capable models:

{
  "agents": {
    "hephaestus": {
      "model": "custom-provider/thinking-model-id",
      "reasoningEffort": "high",
      "allow_non_gpt_model": true
    }
  }
}

Testing

• All existing tests pass
• Tests verify warning toast is shown when allow_non_gpt_model: true

Related Issues

• Feature Request: Support for custom provider in default categories #1071 - Support for custom provider in default categories
• [Feature]: Native OpenRouter provider support #1637 - Native OpenRouter provider support

---

Ultraworked with Sisyphus

[github-actions]

All contributors have signed the CLA. Thank you! ✅
_{Posted by the CLA Assistant Lite bot.}

[cubic-dev-ai]

No issues found across 4 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

_{Auto-approved: The changes safely relax provider restrictions for Hephaestus and implement silent opt-in for non-GPT models as requested, with appropriate test updates and no detected side effects.}

[imwxc]

Closing this PR. Will create an issue to request this feature instead, as it needs more discussion and potentially different implementation approach.