Repository navigation
reasoningToggle() returns {} for @ai-sdk/openai-compatible despite models.dev declaring a toggle (GLM 4.7, GLM 5-turbo) #42793
Description
Description
Problem
GLM 4.7 and GLM 5-turbo (zai-coding-plan) declare reasoning_options: [{"type": "toggle"}] on models.dev, but reasoningToggle() in packages/opencode/src/provider/transform.ts only handles @ai-sdk/alibaba and @ai-sdk/cohere. It returns {} for @ai-sdk/openai-compatible. No reasoning variants appear in the /models menu, so thinking can't be disabled for these models.
Aware of #18598 (closed as intentional). That decision predated the metadata-driven variant system (#36624, #36543, #37519), under which models.dev now declares a toggle for these models, but the new code path can't serve it for this transport. Flagging the inconsistency rather than relitigating the old behavior. If the metadata is wrong instead, that'd be good to know too.
Verification
Against the Z.AI coding plan endpoint, glm-4.7, same prompt:
thinking: {"type": "disabled"}→reasoning_contentempty, correct answer- default → 269 chars of
reasoning_content, correct answer
The API honors the toggle. Only the menu variants are missing.
Suggested fix
case "@ai-sdk/openai-compatible":
return {
none: { thinking: { type: "disabled" } },
high: { thinking: { type: "enabled" } },
}Metadata-gated (only called when models.dev declares a toggle), fix is similar to the MiniMax M3 variants from #38330. Happy to send a PR with test.each coverage if the approach looks right.
Plugins
No response
OpenCode version
1.18.18
Steps to reproduce
- Authenticate zai-coding-plan
- /models → GLM 4.7
- No reasoning variants offered (thinking can't be disabled)
Screenshot and/or share link
No response
Operating System
macOS 15.6.1
Terminal
No response
Activity
[-]easoningToggle() returns {} for @ai-sdk/openai-compatible despite models.dev declaring a toggle (GLM 4.7, GLM 5-turbo)[/-][+]reasoningToggle() returns {} for @ai-sdk/openai-compatible despite models.dev declaring a toggle (GLM 4.7, GLM 5-turbo)[/+]github-actions commentedon Aug 15, 2026
This issue might be a duplicate of or related to existing issues. Please check:
transform.ts), same provider type (@ai-sdk/openai-compatible), GLM model reasoning variant bug — though that one is about wrong variant names (maxvsxhigh) rather than missing toggle supportreasoningToggle()/variants()— same general root cause patternroblennon commentedon Aug 15, 2026
Neither is a duplicate, but #33303 shares a root cause:
GLM-5.2 on OpenAI-compatible providers: TUI variant switcher shows max (invalid reasoning_effort) instead of xhigh #34278 is a different defect — wrong variant names (max vs xhigh) for GLM-5.2 on spec-compliant OpenAI-compatible gateways. Here the endpoint is Z.AI's native API, where
maxis a validreasoning_effort, and the gap is the missing toggle for 4.7/5-turbo.Qwen3.x reasoning models don't show thinking level switcher despite supporting thinking_budget #33303 is the legacy
id.includes("qwen")blanket exclusion, but models.dev now declarestoggle + budget_tokensfor qwen3.x. On a custom OpenAI-compatible provider (their DashScope setup), that toggle funnels into the samereasoningToggle()gap — the openai-compatible case suggested above would likely cover those users too. Direct@ai-sdk/alibabausers already getenableThinkingvariants.fancive commentedon Aug 22, 2026
I can take this. I will first add focused
reasoningVariantscoverage for models.dev-gated toggles, then keep the implementation scoped to the@ai-sdk/openai-compatibleprovider options.fix(models): give toggle-kind models a thinking control in the CLI
thdxr commentedon Oct 11, 2026
Thanks for the report! OpenCode 2.0 is now out and development is focused on V2, so we're closing issues filed against OpenCode 1. Please upgrade (
opencode upgrade --major, details: https://anoma.ly/notes/opencode-2-0), and if this is still relevant on 2.0, reopen this issue or open a new one with your 2.0 version.— from 𝕺𝖕𝖊𝖓𝕮𝖔𝖉𝖊
DXXS commentedon Oct 23, 2026
FWIW: After upgrading from 1.19 to 2.0, when using my MiMo-V2.6-Flash token plan, it is no longer allowing /variants or /thinking or /effort switching!?