Question: should allow_non_gpt_model also bypass Hephaestus registration checks?

[dyyxxhh]

Question

Should agents.hephaestus.allow_non_gpt_model: true also allow Hephaestus to be registered when agents.hephaestus.model is a non-GPT model?

I may be misunderstanding the intended behavior, so I wanted to ask before treating this as a bug.

Context

I saw commit 997db0e05b59a60add35de208c7338158fdf007d, which adds allow_non_gpt_model under agents.hephaestus and passes it into the no-hephaestus-non-gpt hook.

With this config:

{
  "agents": {
    "hephaestus": {
      "model": "volcengine/glm-5.2",
      "mode": "all",
      "allow_non_gpt_model": true
    }
  }
}

I expected Hephaestus - Deep Agent to still appear in the OpenCode agent list, using the configured non-GPT model, with only a warning from the hook.

What I observed

On oh-my-openagent@4.11.1, Hephaestus is still skipped during registration when the configured model is non-GPT.

From reading the bundled code, maybeCreateHephaestusConfig() still checks isHephaestusSupportedModel(...) before returning the agent. As far as I can tell, allow_non_gpt_model is currently only used by the runtime no-hephaestus-non-gpt hook, after an agent/session already exists.

So the option prevents switching away from Hephaestus at runtime, but does not allow a non-GPT Hephaestus to appear in /agent in the first place.

Environment

• oh-my-openagent: 4.11.1
• opencode: 1.17.7
• Hephaestus model: volcengine/glm-5.2
• Config path: agents.hephaestus.allow_non_gpt_model: true

Clarification requested

Is the intended behavior:

1. allow_non_gpt_model should only affect the runtime hook, and Hephaestus registration should remain GPT-only; or
2. allow_non_gpt_model should also bypass the registration-time isHephaestusSupportedModel(...) checks?

If option 2 is intended, I think the registration path may need to read agents.hephaestus.allow_non_gpt_model as well.

[dyyxxhh]

I opened a PR with the implementation I think matches option 2:

#5402

If my understanding is correct, allow_non_gpt_model: true should be threaded through both places:

1. The registration-time maybeCreateHephaestusConfig() model gates, so an explicitly opted-in non-GPT Hephaestus can appear in the agent list.
2. The Hephaestus prompt construction path, so the non-GPT model does not throw while building the prompt. In the PR I made unsupported opted-in models fall back to the generic GPT prompt, while preserving the existing default behavior for users who do not set allow_non_gpt_model.

If the intended behavior is that the option should only affect the runtime hook, then the PR can be closed and this issue is just a clarification request.

[EvangelosMoschou]

Closing my local fix in favor of @dyyxxhh's #5402 which targets the same code paths (registration gate + prompt construction fallback to generic GPT prompt) and is cleaner — their AgentOverrides type already exposes allow_non_gpt_model without a cast. Thanks @dyyxxhh for the implementation and @MoerAI for the pointer.

[WhiteGiverMa]

maybe a warning left and now it makes you must switch to other models manually is what yeongyu want. We can have a good discussion because his obsession with QA is understandable, but sometimes users do want to do it - on the premise of knowing the consequences. We expect it to crash, but why not have a try if miracle? XD

[WhiteGiverMa]

all in all, have a nice talk in the PR qwq