`OPENAI_API_KEY` alone produces silent empty responses

[hunhoon21]

Summary

Setting only OPENAI_API_KEY (without ANTHROPIC_API_KEY or AI Gateway config) results in the chatbot returning empty responses with no error messages. The root cause is that openclaw onboard --auth-choice openai-api-key saves the key to ~/.openclaw/.env but does not set agents.defaults.model in the config, so the gateway defaults to anthropic/claude-opus-4-5 — which fails silently with no Anthropic key.

Steps to Reproduce

1. Deploy moltworker with only OPENAI_API_KEY set (no ANTHROPIC_API_KEY, no AI Gateway secrets)
2. Open the Control UI and send a message
3. The assistant responds with an empty message — no error, no indication of what went wrong

Root Cause

In start-openclaw.sh, the onboard step runs:

openclaw onboard --non-interactive --accept-risk \
    --auth-choice openai-api-key --openai-api-key $OPENAI_API_KEY \
    ...

This correctly saves the API key to ~/.openclaw/.env, but the resulting openclaw.json has no agents.defaults.model entry. The gateway then falls back to anthropic/claude-opus-4-5 as the default model, which requires an Anthropic API key that isn't configured.

The CF_AI_GATEWAY_MODEL override block in the config patch handles AI Gateway models, but there's no equivalent fallback for direct OPENAI_API_KEY usage.

Expected Behavior

Setting OPENAI_API_KEY alone should work out of the box. The config patch should detect this case and set the default model to openai/gpt-4o (or another OpenAI model).

Proposed Fix

Add a fallback block in the start-openclaw.sh config patch (after the CF_AI_GATEWAY_MODEL block) that detects when OPENAI_API_KEY is set without an AI Gateway model override and sets the default model accordingly:

if (process.env.OPENAI_API_KEY && !process.env.CF_AI_GATEWAY_MODEL) {
    const currentModel = config.agents?.defaults?.model?.primary || '';
    if (!currentModel || currentModel.startsWith('anthropic/')) {
        config.agents = config.agents || {};
        config.agents.defaults = config.agents.defaults || {};
        config.agents.defaults.model = { primary: 'openai/gpt-4o' };
    }
}

Additionally, the README should be updated to list OPENAI_API_KEY as a standalone alternative to ANTHROPIC_API_KEY in the Requirements and Quick Start sections.

Environment

• moltworker deployed on Cloudflare Sandbox
• openclaw@2026.2.3 (container image)
• Only OPENAI_API_KEY set as AI provider secret

AI Disclosure

This issue was investigated and drafted with assistance from Claude Code. The bug was discovered through manual testing and confirmed by inspecting /debug/container-config and /debug/processes?logs=true outputs. The proposed fix has been tested and verified on a live deployment.

[pablomichelettii]

i've also had this error. I've solved it in similar fashion. But i have a problem with the ai gateway that i can't make it work with openai. However, the main issue of debugging this things is that the chat fails silently.

I suggest to implement the readme with other info about this