GLM-5.2 on OpenAI-compatible providers: TUI variant switcher shows max (invalid reasoning_effort) instead of xhigh

[bogerman1]

Description

GLM-5.2 routed through an OpenAI-compatible provider ("api": "openai") exposes the wrong reasoning_effort variants in the TUI. The variant switcher shows high and max, but max is not part of the OpenAI reasoning_effort enum (none/minimal/low/medium/high/xhigh). Selecting max causes a 400 from spec-compliant upstreams (e.g. hyper, Cloudflare AI Gateway). The actually-selectable top tier xhigh is never surfaced.

Root cause is in packages/opencode/src/provider/transform.ts — the glm-5.2 + @ai-sdk/openai-compatible branch returns { high, max } instead of { high, xhigh }. The sibling OpenRouter branch already does the right thing (exposes xhigh, gateway translates to GLM's native max).

Plugins

None.

OpenCode version

1.17.11

Steps to reproduce

1. Configure an OpenAI-compatible provider pointing at a GLM-5.2 backend:

   {
     "provider": {
       "hyper": {
         "id": "hyper", "name": "Hyper", "api": "openai",
         "options": { "apiKey": "...", "baseURL": "https://hyper.charm.land/v1" },
         "models": {
           "glm-5.2": {
             "name": "GLM-5.2", "reasoning": true,
             "interleaved": { "field": "reasoning_content" },
             "tool_call": true,
             "limit": { "context": 1000000, "output": 128000 }
           }
         }
       }
     },
     "model": "hyper/glm-5.2@xhigh"
   }

2. Launch opencode, open the TUI.

3. Open the model variant switcher — only high and max are listed; xhigh is missing.

4. Select max and send a prompt — upstream returns 400 (reasoning_effort: "max" rejected).

Screenshot and/or share link

N/A (config + 400, no share link).

Operating System

macOS 26 (arm64)

Terminal

Ghostty

Additional context

The OpenRouter branch in the same function already handles this correctly:

if (glm52 && model.api.npm === "@openrouter/ai-sdk-provider") {
  // OpenRouter maps xhigh to GLM-5.2's native max effort.
  return { high: { reasoning: { effort: "high" } }, xhigh: { reasoning: { effort: "xhigh" } } }
}

The @ai-sdk/anthropic branch is also correct as-is (Anthropic's effort field accepts max natively). Only the OpenAI-compatible branch needs fixing.