openclaw - 💡(How to fix) Fix [Bug]: Ollama Cloud think modes don't match mapped choices in /think [2 comments, 2 participants]

Description

When using Ollama cloud models via a base URL connection, OpenClaw's /think command offers a hardcoded set of thinking levels that don't match what the upstream Ollama API accepts per model. Specifically, it offers minimal (which isn't a native Ollama thinking level) and omits max (which some models support).

Background

Ollama's thinking support varies by model:

• Most thinking models accept true/false (on/off)
• GPT-OSS accepts only low, medium, high — and ignores true/false entirely (Ollama docs)
• Models like deepseek-v4-pro also support max (ollama/ollama#15952)

The Ollama cloud API validates think against high, medium, low, max, true, false at the API layer, but not all models meaningfully support all levels.

Interestingly, the API validation layer exposes max even for GPT-OSS, despite the docs claiming it only supports low/medium/high:

$ docker exec -it ollama ollama run gpt-oss:120b:cloud
Connecting to 'gpt-oss:120b:cloud' on 'ollama.com' ⚡
>>> /set think test
Set 'think' mode to 'test'.
>>> test
error: 400 Bad Request: invalid think value: "test" (must be "high", "medium", "low", "max", true, or false)

What OpenClaw offers via /think

• off, minimal, low, medium, high

Problems:

1. minimal is not a native Ollama thinking level. It appears to work, so OpenClaw is presumably mapping it to something else before sending the request — but this is opaque and confusing. It looks like a first-class option the API doesn't recognise.
2. max is missing. Models that support max thinking have no way to enable it through OpenClaw.
3. The list is hardcoded, not per-model. GPT-OSS only supports low/medium/high, while other models have different sets of valid levels. OpenClaw presents the same choices regardless of which model is active.

Expected behavior

OpenClaw should discover the thinking modes supported by each model (or at least per provider) rather than using a static hardcoded list. The available /think options should reflect what the upstream API actually accepts for the active model.

Reproduction

1. Configure OpenClaw with an Ollama base URL (not API key) pointing to a local Ollama instance with cloud model access
2. Switch to a cloud model: /model ollama/gpt-oss:120b:cloud
3. Run /think
4. Observe that minimal is offered (not a valid Ollama thinking level) and max is missing (valid for some models)

Environment

• OpenClaw 2026.5.4 (325df3e)
• OpenClaw connected to Ollama via base URL (not API key)
• Ollama cloud model (:cloud suffix)

Issue submitted by civil's clanker