ollama - 💡(How to fix) Fix launch/codex: map unsupported xhigh reasoning effort for Ollama models

Summary

ollama launch codex can expose Codex reasoning metadata for Ollama models that report the thinking capability. If the user's Codex config has model_reasoning_effort = "xhigh", Codex sends xhigh to Ollama, but Ollama's Responses API currently accepts high, medium, low, max, or none.

That causes a request failure for thinking-capable Ollama models instead of cleanly launching Codex.

Repro

With ~/.codex/config.toml containing:

model_reasoning_effort = "xhigh"

Run:

ollama launch codex --model deepseek-v4-pro:cloud

Then send a prompt such as:

hi

Observed error:

invalid reasoning value: "xhigh" (must be "high", "medium", "low", "max", or "none")

Why this happens

Codex supports xhigh for some OpenAI models, and users may have that as their global Codex preference. Ollama's API uses a different accepted reasoning-effort enum. When the Ollama-generated Codex model catalog advertises reasoning support, Codex applies the global xhigh setting to the Ollama-backed request.

Expected behavior

ollama launch codex should avoid sending unsupported reasoning effort values to Ollama.

Possible approaches:

• Downshift xhigh to high for the generated ollama-launch Codex profile.
• Map Codex reasoning levels to Ollama-supported values in the generated metadata/config.
• Omit reasoning capability metadata until the mapping is explicit.