Summary

On openclaw 2026.3.7, openai-codex/gpt-5.4 can resolve to a model object with api: undefined when the user has a custom models.providers.openai-codex block in ~/.openclaw/openclaw.json that defines the model but omits api.

In my case this caused the gateway to crash-loop on startup when cron replayed missed isolated jobs, with:

Unhandled promise rejection: Error: No API provider registered for api: undefined

Environment

• OpenClaw: 2026.3.7
• Install type: non-Docker user install
• Gateway launched via user systemd service
• Provider: openai-codex
• Model: gpt-5.4

Minimal config shape that triggers it

This custom provider block was present in ~/.openclaw/openclaw.json:

{
  "models": {
    "providers": {
      "openai-codex": {
        "baseUrl": "https://chatgpt.com/backend-api",
        "models": [
          {
            "id": "gpt-5.3-codex",
            "name": "gpt-5.3-codex"
          },
          {
            "id": "gpt-5.4",
            "name": "gpt-5.4"
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "openai-codex/gpt-5.4"
      }
    }
  }
}

Note that api is omitted from the custom provider block.

Repro

1. Configure a custom models.providers.openai-codex block like the above, without api.
2. Set agents.defaults.model.primary to openai-codex/gpt-5.4.
3. Run an isolated cron agentTurn job, or restart the gateway so missed cron jobs replay.
4. The gateway eventually crashes with:

No API provider registered for api: undefined

Expected

One of these should happen:

• the built-in openai-codex/gpt-5.4 metadata should still win or be merged, preserving api: openai-codex-responses
• or the config should be rejected/validated with a clear error because the custom provider/model definition is incomplete

Actual

The custom models.providers.openai-codex definition appears to shadow the built-in gpt-5.4 model metadata. That leaves gpt-5.4 with no resolved API, which later crashes in the streaming path with api: undefined.

Likely root cause

From local inspection of the built bundle in 2026.3.7:

• gpt-5.4 is known to the built-in catalog / forward-compat path for openai-codex
• but buildInlineProviderModels() creates inline models with:
  • api: model.api ?? entry.api
• if the custom provider block omits both, the inline model has no api
• and that custom provider/model path seems to take precedence over the built-in forward-compat model for openai-codex/gpt-5.4

So this looks like a precedence/merge issue between custom provider config and built-in model metadata.

Local workaround

Adding this fixed it immediately for me:

"models": {
  "providers": {
    "openai-codex": {
      "api": "openai-codex-responses",
      "baseUrl": "https://chatgpt.com/backend-api",
      "models": [
        { "id": "gpt-5.4", "name": "gpt-5.4" }
      ]
    }
  }
}

After adding api: "openai-codex-responses", the gateway came up cleanly on openai-codex/gpt-5.4 and stopped crashing.

Suggestion

It would help if OpenClaw either:

• merged built-in model metadata into custom provider models when provider/model matches a known built-in model, or
• validated models.providers.* entries so required transport fields like api cannot be omitted silently for providers that need them