fix(provider): support openai-compatible reasoning toggle

[fancive]

Intent

修复 anomalyco/opencode issue #42793：当 models.dev 为 @ai-sdk/openai-compatible 模型声明 reasoning toggle 时，ProviderTransform.reasoningVariants 应生成 none/high 变体，分别发送 thinking.type=disabled/enabled，使 GLM 4.7 与 GLM 5 Turbo 等兼容端点可关闭或启用 thinking。实现应严格限制在 models.dev 声明 toggle 且 npm 为 @ai-sdk/openai-compatible 的路径，不改变 effort、budget 或其他 provider 行为；添加可执行行为回归测试。先完成充分本地测试，再将单一小提交推送到 fancive fork 并创建以 dev 为 base 的 GitHub PR，关联 issue #42793。

What Changed

• ProviderTransform.reasoningToggle now also recognizes models with npm @ai-sdk/openai-compatible (previously only @ai-sdk/cohere), so a reasoning toggle declared in models.dev produces none/high variants that send thinking.type: "disabled"/"enabled" for compatible endpoints such as GLM 4.7 and GLM 5 Turbo.
• Added a regression test asserting the openai-compatible toggle mapping in ProviderTransform.reasoningVariants.

Risk Assessment

⚠️ Medium: The core fix is small, correctly implements the intended GLM 4.7/GLM 5 Turbo thinking toggle (verified that @ai-sdk/openai-compatible 2.0.41 passes unknown provider options like thinking through into the request body), and ships a proper behavioral regression test, but it also silently shadows deliberate variants() fallback behavior for minimax-m3 (nvidia/lilac chat_template_kwargs) and toggle-only glm-5.2 gateway models, which the author should confirm is acceptable.

Testing

对 issue #42793 的修复做了定向验证:新增回归测试在 base commit 上失败、修复后通过;transform 全文件 412 测试通过;用实时 models.dev 数据端到端复现 GLM 4.7 / GLM 5 Turbo / GLM 5(openai-compatible + toggle)生成 none/high 变体,并经真实 AI SDK 捕获到发送到端点的请求体分别携带 thinking.type=disabled/enabled,同时确认 effort 声明模型(glm-5.2)仍走 reasoningEffort 变体、其它 provider 行为不变,满足用户意图的全部约束。

Evidence: 端到端证据:GLM 4.7 toggle 变体生成与 wire 请求体(真实 models.dev 数据 + 真实 @ai-sdk/openai-compatible SDK)

{ "source": "models.dev/api.json (fetched live)", "provider": "zai (ZhipuAI)", "npm": "@ai-sdk/openai-compatible", "model": "glm-4.7", "variants": { "none": { "thinking": { "type": "disabled" } }, "high": { "thinking": { "type": "enabled" } } }, "wire": [ { "variant": "none", "wireBody": { "thinking": { "type": "disabled" }, "model": "glm-4.7" } }, { "variant": "high", "wireBody": { "thinking": { "type": "enabled" }, "model": "glm-4.7" } } ], "effortModelCheck": { "model": "glm-5.2", "variants": { "high": { "reasoningEffort": "high" }, "max": { "reasoningEffort": "max" } } } }

{
"source": "models.dev/api.json (fetched live)",
"provider": "zai (ZhipuAI)",
"npm": "@ai-sdk/openai-compatible",
"model": "glm-4.7",
"variants": {
"none": {
"thinking": {
"type": "disabled"
}
},
"high": {
"thinking": {
"type": "enabled"
}
}
},
"wire": [
{
"variant": "none",
"providerOptions": {
"zai": {
"thinking": {
"type": "disabled"
}
}
},
"wireBody": {
"thinking": {
"type": "disabled"
},
"model": "glm-4.7"
}
},
{
"variant": "high",
"providerOptions": {
"zai": {
"thinking": {
"type": "enabled"
}
}
},
"wireBody": {
"thinking": {
"type": "enabled"
},
"model": "glm-4.7"
}
}
],
"effortModelCheck": {
"model": "glm-5.2",
"variants": {
"high": {
"reasoningEffort": "high"
},
"max": {
"reasoningEffort": "max"
}
}
}
}

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

⚠️ **Review** - 1 warning

• ⚠️packages/opencode/src/provider/transform.ts:1711 - Adding @ai-sdk/openai-compatible to reasoningToggle makes reasoningVariants return a non-undefined value for every openai-compatible model that models.dev declares a toggle for, which shadows the variants() fallback special cases (reasoningVariants ?? variants in provider.ts:1289). Concrete regression: nvidia minimaxai/minimax-m3 and lilac minimaxai/minimax-m3 (models.dev declares [{type:"toggle"}], npm @ai-sdk/openai-compatible) previously fell back to the deliberate minimax-m3 branch in variants() (transform.ts:736-749, added by fix(provider): correct MiniMax M3 thinking variants #38330) producing {none: {chat_template_kwargs: {thinking_mode: "disabled"}}, thinking: {chat_template_kwargs: {thinking_mode: "enabled"}}} for nvidia/lilac, and {none: {thinking: {type: "disabled"}}, thinking: {thinking: {type: "adaptive"}}} for others. After this change they get {none: {thinking: {type: "disabled"}}, high: {thinking: {type: "enabled"}}} — the variant id 'thinking' disappears, and for nvidia/lilac the endpoint-specific chat_template_kwargs control is bypassed; if the NIM endpoint ignores the generic 'thinking' field (the premise of fix(provider): correct MiniMax M3 thinking variants #38330), selecting the high variant silently fails to enable thinking (wrong result without error). Similarly, toggle-only glm-5.2 gateway models (e.g. crossmodel z-ai/glm-5.2, nvidia z-ai/glm-5.2) lose the {high/max: reasoningEffort} variants from the glm52 fallback (transform.ts:758-761), changing user-visible options beyond the GLM 4.7/5 Turbo scope. This conflicts with the intent constraint '不改变 effort、budget 或其他 provider 行为'. Recommended boundary: in reasoningVariants/reasoningToggle, keep the existing special-cased models (minimax-m3, glm-5.2) on their prior fallback path (early return for those ids) so the fix only activates where no deliberate fallback behavior exists.

✅ **Test** - passed

✅ No issues found.

• bun test test/provider/transform.test.ts -t reasoningVariants — 55 个变体测试全部通过,含新增 converts openai-compatible toggle options
• 回归验证:临时将 src/provider/transform.ts 换回 base commit e00890c 版本后运行新测试 → 失败(返回 undefined,证明修复前无法生成变体);恢复修复版本后 → 通过
• bun test test/provider/transform.test.ts — 全文件 412 测试全部通过,781 个断言,无回归
• 端到端脚本:实时拉取 models.dev/api.json,zai(ZhipuAI,npm=@ai-sdk/openai-compatible)provider 的 glm-4.7 经 fromModelsDevProvider 生成 variants none/high(thinking disabled/enabled)
• 端到端脚本:变体经 ProviderTransform.providerOptions 后送入真实 @ai-sdk/openai-compatible SDK(generateText + 捕获 fetch),断言 wire JSON body 含 thinking: {type: "disabled"} / {type: "enabled"}
• 范围约束:zai 的 glm-5.2(effort 声明)仍生成 reasoningEffort high/max 变体,effort 路径未被 toggle 路径影响;glm-5-turbo、glm-5 同样正确生成 none/high 变体
• diff 审查:transform.ts 仅 1 行改动(cohere 分支追加 || model.api.npm === "@ai-sdk/openai-compatible"),effort 优先分支与其它 provider 行为不变

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.

[github-actions]

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

• PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

This pull request has been automatically closed because it was not updated to meet our contributing guidelines within the 2-hour window.

Feel free to open a new pull request that follows our guidelines.