[Bug] Refinement fails if current model does not support no thinking

[moutazhaq]

Affected area

Coding agent and CLI

What happened?

Refinement fails if the current model does not support no thinking mode.

Steps to reproduce

1. Run prime-agent.
2. Select a model that doesn't support no thinking mode (for example, zai/glm-5.3).
3. Trigger a refinement.
4. Observe error: Error: Refinement failed: Refinement failed: 400 This model always engages in thinking and cannot be disabled; please use low, high, or max

Expected behavior

The refinement should succeed.

Prime Agent version

849c921

Environment

Arch Linux (CachyOS), Wezterm

Additional context

No response

[moutazhaq]

Trying to use /btw has the same issue.

[aljazceru]

i have a PR that addresses this #1565

[aljazceru]

Root cause and a fix are up in #1565.

The z.ai coding endpoint (api.z.ai/api/coding/paas/v4) rejects enable_thinking: false for effort-capable GLM models (5.2+). Live-probed against the endpoint:

payload	glm-4.7	glm-5-turbo	glm-5.2	glm-5.3
enable_thinking: false	200, no thinking	200, no thinking	400	400
reasoning_effort: "none"	—	—	200, no thinking	200, no thinking
no thinking params	thinks	thinks	thinks	thinks

"Thinking off" for 5.2/5.3 must be expressed as reasoning_effort: "none", not a disable toggle — matching z.ai's docs ("none or minimal indicate that the model stops thinking").

Both failing flows send a disable: /btw hardcodes thinkingLevel: "off" (side-question.ts), and refinement omits reasoning, which the zai compat branch translated to enable_thinking: false (openai-completions.ts). The PR maps off → reasoning_effort: "none" via a thinkingLevelMap for direct z.ai models, keeps toggle-only routes (glm-4.7/5-turbo, Prime Inference z-ai) unchanged, and makes background completions request an explicit clamped level. Same fix also resolves #1456

[inxeoz]

Facing same issue with Muse Spark models via Console/opencode:

• muse-spark-1.2-contributor-free
• opencode/muse-spark-1.2-contributor-free
• muse-spark-1.2 / meta/muse-spark-*

Repro: /model muse-spark-1.2-contributor-free -> /refine fails with reasoning.effort does not support "none" (Console).

i am using custom extension patch which lets me override thinking effort https://github.com/inxeoz/prime-agent-ext/blob/master/fix-muse-spark-off.ts