#openrouter/auto thinking issue

1 messages · Page 1 of 1 (latest)

rich tiger
#

I'm using openrouter/auto but keep getting this thinking warning:

[agent/embedded] unsupported thinking level for openrouter/openrouter/auto; retrying with minimal

I believe this is impacting latency as well as quality of my LLM calls

How can I fix it?

tidal breachBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> âś… Mark Solution

shrewd valveBOT
#

Let me verify the exact behavior in docs/source so I don’t hand you a made-up config key.

#

Short version: yes, that retry can add latency.

What’s happening:

  • openrouter/auto is a moving target, so thinking support is not stable.
  • OpenClaw sees a thinking level, OpenRouter rejects it for the routed backend, then OpenClaw retries with minimal.

Best fix:

  1. Paste openclaw --version first.
  2. If you want stable thinking/quality, stop using openrouter/auto and switch to a concrete OpenRouter model.
  3. If you want to stay on openrouter/auto, turn thinking off for that session/model:
    • chat: /think off
    • or config:
    {
  agents: {
    defaults: {
# 
         models: {
           "openrouter/auto": {
             params: { thinking: "off" },
           },
         },
       },
     },
   }

You can also use agents.defaults.thinkingDefault: "off" if you want it globally.

I checked the docs/source, and OpenClaw’s OpenRouter docs say supported non-auto routes get reasoning injection, while openrouter/auto should skip it. So if you’re on a recent version and still seeing this a lot, it may be version-specific.

Docs:

rich tiger
#

OpenClaw 2026.4.21

shrewd valveBOT
#

Let me check whether that looks like known-by-code behavior on 2026.4.21 or more like a rough edge.

#

Thanks, that helps. On 2026.4.21, I’d treat this as a real rough edge, not just “normal tuning.”

Why:

  • OpenClaw’s docs/source say openrouter/auto should skip reasoning injection.
  • But openrouter/auto is still a router, so the actual backend can change per call.
  • So yes, the retry can hurt latency, and auto also hurts quality consistency even when it doesn’t warn.

What I’d do:

  1. In that thread, send /think and check whether a session thinking override is stuck.
  2. If you want to stay on openrouter/auto, send /think off.
  3. If you want stable reasoning behavior, stop using openrouter/auto and switch to a concrete OpenRouter model.

If you want the config version instead of chat commands:

# 
{
  agents: {
    defaults: {
      models: {
        "openrouter/auto": {
          params: { thinking: "off" },
        },
      },
    },
  },
}

So my blunt recommendation is:

  • Convenience: keep openrouter/auto, but turn thinking off