Custom provider fails only when reasoning is enabled

Symptoms

• Basic chat works.
• The same provider fails after enabling reasoning or thinking features.
• Errors may appear only for certain models or only after changing defaults.
• Tool calling may also become unstable once reasoning is enabled.

Cause

“Reasoning-capable model” and “this endpoint accepts reasoning-control fields” are not the same thing.

A custom OpenAI-compatible endpoint may accept normal chat requests but reject or mis-handle fields such as:

• reasoning_effort,
• provider-specific thinking controls,
• or translated reasoning payloads passed through a proxy layer.

This is why a model can be genuinely reasoning-capable while the compatibility layer in front of it is not.

Fix

1) Turn reasoning off for the compatibility test

When debugging, first prove the endpoint can survive a plain run:

{
  models: {
    providers: {
      myprovider: {
        models: [
          {
            id: "my-model",
            reasoning: false,
          },
        ],
      },
    },
  },
}

Then retry with a fresh session.

2) Check whether the failure appears only after changing thinking defaults

Look for:

• global thinking defaults,
• per-model reasoning flags,
• or a recent configuration change that enabled reasoning-like behavior.

If the problem started only after that change, your endpoint likely rejects the reasoning-control layer, not the model itself.

3) Treat the endpoint as plain chat until proven otherwise

For custom compatibility layers, the safe default is:

• reasoning: false during initial integration,
• then re-enable only after proving the specific endpoint contract supports it.

4) Distinguish model capability from transport compatibility

The right mental model is:

• the model may support reasoning,
• but the proxy, relay, or OpenAI-compatible wrapper may not support the control fields OpenClaw uses.

That is a transport limitation, not proof that the underlying model is weak or broken.

Verify

You have confirmed the diagnosis if:

• the provider works with reasoning off,
• fails only when reasoning is enabled,
• and becomes stable again after reverting to plain chat mode.

Related

• Runtime compatibility overview: /guides/openclaw-relay-and-api-proxy-troubleshooting
• Config debugging guide: /guides/openclaw-configuration
• If the endpoint also breaks on other runtime fields: /troubleshooting/solutions/openai-compatible-endpoint-rejects-stream-or-store