openclaw - 💡(How to fix) Fix [Feature]: Support reasoning_effort parameter for Grok 4.3 model [1 pull requests]

Summary

OpenClaw currently does not expose the reasoning_effort parameter when using Grok 4.3 models. According to the official xAI documentation, grok-4.3 supports controlling the amount of reasoning the model performs before responding via the reasoning_effort parameter.

Current behavior: All requests to grok-4.3 (and similar reasoning-capable models) are sent without the reasoning parameter, which defaults to "low".

Requested behavior: Allow users to set reasoning_effort to one of the following values:

• "none" — Disable reasoning entirely (fastest)
• "low" — Default, balanced for agentic/tool use
• "medium" — More thinking for complex tasks
• "high" — Maximum reasoning depth for hard problems

Problem to solve

• Many users want the ability to trade off speed vs. quality depending on the task.
• Complex analysis, math, multi-step logic, and coding tasks benefit significantly from "medium" or "high" reasoning effort.
• Without this parameter, users are stuck with the default behavior even when they want deeper reasoning.

Proposed solution

Support reasoning_effort as a model-specific option, for example:

   {
     "models": {
       "xai/grok-4.3": {
         "params": {
           "reasoningEffort": "high"
         }
       }
     }
   }

Allow per-agent overrides so different agents can use different reasoning levels:

   {
     "agents": {
       "researcher": {
         "model": {
           "primary": "grok-4.3",
           "reasoningEffort": "high"
         }
       }
     }
   }

Alternatives considered

No response

Impact

Provide users with the ability to take advantage of higher reasoning capabilities of Grok 4.3

Evidence/examples

No response

Additional information

https://docs.x.ai/developers/model-capabilities/text/reasoning