fix(cf-ai-gateway): route provider options through openaiCompatible key (#24432)

packages/opencode/src/provider/transform.ts

@@ -41,6 +41,13 @@ function sdkKey(npm: string): string | undefined {
       return "gateway"
     case "@openrouter/ai-sdk-provider":
       return "openrouter"
+    case "ai-gateway-provider":
+      // ai-gateway-provider/unified wraps createOpenAICompatible({ name: "Unified" }),
+      // and @ai-sdk/openai-compatible parses compatibleOptions from one of
+      // "openai-compatible" / "openaiCompatible" / "Unified" / "unified". The
+      // "openai-compatible" key emits a deprecation warning at runtime, so we
+      // pick the camelCase form the SDK now treats as canonical.
+      return "openaiCompatible"
   }
   return undefined
 }
@@ -427,6 +434,36 @@ export function topK(model: Provider.Model) {
 const WIDELY_SUPPORTED_EFFORTS = ["low", "medium", "high"]
 const OPENAI_EFFORTS = ["none", "minimal", ...WIDELY_SUPPORTED_EFFORTS, "xhigh"]
 
+// OpenAI rolled out the `none` reasoning_effort tier on this date (Responses API).
+// Models released before it 400 on `reasoning_effort: "none"`, so we only expose
+// it as a variant for models new enough to accept it.
+const OPENAI_NONE_EFFORT_RELEASE_DATE = "2025-11-13"
+
+// OpenAI rolled out the `xhigh` reasoning_effort tier on this date. Same reasoning.
+const OPENAI_XHIGH_EFFORT_RELEASE_DATE = "2025-12-04"
+
+// Matches members of the gpt-5 family across the id formats we encounter:
+//   "gpt-5", "gpt-5-nano", "gpt-5.4", "openai/gpt-5.4-codex".
+// Anchored to start-of-string or "/" so it doesn't false-match "gpt-50" or "gpt-5o".
+const GPT5_FAMILY_RE = /(?:^|\/)gpt-5(?:[.-]|$)/
+
+// Computes the reasoning_effort tiers an OpenAI (or OpenAI-compatible upstream
+// routed through it, e.g. cf-ai-gateway) model exposes. Returns null for models
+// with no tunable effort knob (gpt-5-pro). Effort order: weakest → strongest.
+function openaiReasoningEfforts(apiId: string, releaseDate: string): string[] | null {
+  const id = apiId.toLowerCase()
+  if (id === "gpt-5-pro" || id === "openai/gpt-5-pro") return null
+  if (id.includes("codex")) {
+    if (id.includes("5.2") || id.includes("5.3")) return [...WIDELY_SUPPORTED_EFFORTS, "xhigh"]
+    return [...WIDELY_SUPPORTED_EFFORTS]
+  }
+  const efforts = [...WIDELY_SUPPORTED_EFFORTS]
+  if (GPT5_FAMILY_RE.test(id)) efforts.unshift("minimal")
+  if (releaseDate >= OPENAI_NONE_EFFORT_RELEASE_DATE) efforts.unshift("none")
+  if (releaseDate >= OPENAI_XHIGH_EFFORT_RELEASE_DATE) efforts.push("xhigh")
+  return efforts
+}
+
 function anthropicAdaptiveEfforts(apiId: string): string[] | null {
   if (["opus-4-7", "opus-4.7"].some((v) => apiId.includes(v))) {
     return ["low", "medium", "high", "xhigh", "max"]
@@ -476,6 +513,21 @@ export function variants(model: Provider.Model): Record<string, Record<string, a
       if (!model.id.includes("gpt") && !model.id.includes("gemini-3") && !model.id.includes("claude")) return {}
       return Object.fromEntries(OPENAI_EFFORTS.map((effort) => [effort, { reasoning: { effort } }]))
 
+    case "ai-gateway-provider": {
+      // Cloudflare AI Gateway routes every upstream through its OpenAI-compatible
+      // /v1/compat endpoint, so the body is always OAI-shaped. The gateway
+      // translates `reasoning_effort` to the upstream provider's native control
+      // (e.g. Anthropic thinking budgets) when needed. Variants therefore stay
+      // OAI-style for all upstreams, with an extended effort set for OpenAI
+      // models that support it.
+      if (model.api.id.startsWith("openai/")) {
+        const efforts = openaiReasoningEfforts(model.api.id, model.release_date)
+        if (!efforts) return {}
+        return Object.fromEntries(efforts.map((effort) => [effort, { reasoningEffort: effort }]))
+      }
+      return Object.fromEntries(WIDELY_SUPPORTED_EFFORTS.map((effort) => [effort, { reasoningEffort: effort }]))
+    }
+
     case "@ai-sdk/gateway":
       if (model.id.includes("anthropic")) {
         if (adaptiveEfforts) {
@@ -595,28 +647,12 @@ export function variants(model: Provider.Model): Record<string, Record<string, a
           },
         ]),
       )
-    case "@ai-sdk/openai":
+    case "@ai-sdk/openai": {
       // https://v5.ai-sdk.dev/providers/ai-sdk-providers/openai
-      if (id === "gpt-5-pro") return {}
-      const openaiEfforts = iife(() => {
-        if (id.includes("codex")) {
-          if (id.includes("5.2") || id.includes("5.3")) return [...WIDELY_SUPPORTED_EFFORTS, "xhigh"]
-          return WIDELY_SUPPORTED_EFFORTS
-        }
-        const arr = [...WIDELY_SUPPORTED_EFFORTS]
-        if (id.includes("gpt-5-") || id === "gpt-5") {
-          arr.unshift("minimal")
-        }
-        if (model.release_date >= "2025-11-13") {
-          arr.unshift("none")
-        }
-        if (model.release_date >= "2025-12-04") {
-          arr.push("xhigh")
-        }
-        return arr
-      })
+      const efforts = openaiReasoningEfforts(model.api.id, model.release_date)
+      if (!efforts) return {}
       return Object.fromEntries(
-        openaiEfforts.map((effort) => [
+        efforts.map((effort) => [
           effort,
           {
             reasoningEffort: effort,
@@ -625,6 +661,7 @@ export function variants(model: Provider.Model): Record<string, Record<string, a
           },
         ]),
       )
+    }
 
     case "@ai-sdk/anthropic":
     // https://v5.ai-sdk.dev/providers/ai-sdk-providers/anthropic

packages/opencode/test/provider/cf-ai-gateway-e2e.test.ts

@@ -0,0 +1,110 @@
+// End-to-end regression test for opencode#24432.
+//
+// Routes through the actual ai-gateway-provider + @ai-sdk/openai-compatible
+// chain that provider.ts:811 builds at runtime, with only the network boundary
+// stubbed. Asserts that `reasoning_effort` (and other provider options the
+// transform emits) actually land in the body Cloudflare AI Gateway forwards
+// upstream — which is the only place the bug was observable.
+
+import { afterEach, beforeEach, describe, expect, test } from "bun:test"
+import { generateText } from "ai"
+import { createAiGateway } from "ai-gateway-provider"
+import { createUnified } from "ai-gateway-provider/providers/unified"
+import { ProviderTransform } from "@/provider/transform"
+
+type Captured = { url: string; outerBody: any }
+
+const realFetch = globalThis.fetch
+let captured: Captured | null = null
+
+beforeEach(() => {
+  captured = null
+  globalThis.fetch = (async (input: any, init?: any) => {
+    const url = typeof input === "string" ? input : input instanceof URL ? input.toString() : input.url
+    if (url.startsWith("https://gateway.ai.cloudflare.com/")) {
+      const bodyText = init?.body ?? ""
+      captured = { url, outerBody: bodyText ? JSON.parse(bodyText) : null }
+      return new Response(
+        JSON.stringify({
+          id: "chatcmpl-test",
+          object: "chat.completion",
+          created: 0,
+          model: "openai/gpt-5.4",
+          choices: [{ index: 0, message: { role: "assistant", content: "ok" }, finish_reason: "stop" }],
+          usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
+        }),
+        { status: 200, headers: { "Content-Type": "application/json" } },
+      )
+    }
+    return realFetch(input, init)
+  }) as typeof fetch
+})
+
+afterEach(() => {
+  globalThis.fetch = realFetch
+})
+
+const cfModel = (apiId: string, releaseDate = "2026-03-05"): any => ({
+  id: `cloudflare-ai-gateway/${apiId}`,
+  providerID: "cloudflare-ai-gateway",
+  api: { id: apiId, url: "https://gateway.ai.cloudflare.com/v1/compat", npm: "ai-gateway-provider" },
+  capabilities: {
+    reasoning: true,
+    temperature: false,
+    attachment: true,
+    toolcall: true,
+    input: { text: true, audio: false, image: true, video: false, pdf: true },
+    output: { text: true, audio: false, image: false, video: false, pdf: false },
+    interleaved: false,
+  },
+  cost: { input: 1, output: 1, cache: { read: 0, write: 0 } },
+  limit: { context: 1_000_000, output: 128_000 },
+  status: "active",
+  options: {},
+  headers: {},
+  release_date: releaseDate,
+})
+
+async function callThroughGateway(apiId: string, providerOptions: Record<string, any>) {
+  const aigateway = createAiGateway({ accountId: "test", gateway: "test", apiKey: "test" })
+  const unified = createUnified()
+  await generateText({ model: aigateway(unified(apiId)), prompt: "hi", providerOptions })
+  // ai-gateway-provider sends an array; each entry's `query` is the upstream body.
+  return captured?.outerBody?.[0]?.query as Record<string, any> | undefined
+}
+
+describe("cf-ai-gateway end-to-end (regression: #24432)", () => {
+  test("ProviderTransform.providerOptions output puts reasoning_effort on the wire", async () => {
+    // The full chain the runtime exercises:
+    //   transform.providerOptions() → openaiCompatible key
+    //   → @ai-sdk/openai-compatible reads it as compatibleOptions
+    //   → emits body.reasoning_effort
+    //   → ai-gateway-provider wraps the body and forwards to gateway.ai.cloudflare.com
+    const opts = ProviderTransform.providerOptions(cfModel("openai/gpt-5.4"), { reasoningEffort: "xhigh" })
+    expect(opts).toEqual({ openaiCompatible: { reasoningEffort: "xhigh" } })
+
+    const upstream = await callThroughGateway("openai/gpt-5.4", opts)
+    expect(upstream?.reasoning_effort).toBe("xhigh")
+  })
+
+  test("variants() output for openai/gpt-5.4 lands xhigh on the wire", async () => {
+    // The other half of the bug: workflow `variant: xhigh` flows through variants()
+    // and must reach the wire. variants() returns the providerOptions payload
+    // unwrapped; providerOptions() wraps it under the SDK key.
+    const variants = ProviderTransform.variants(cfModel("openai/gpt-5.4"))
+    expect(variants.xhigh).toEqual({ reasoningEffort: "xhigh" })
+
+    const opts = ProviderTransform.providerOptions(cfModel("openai/gpt-5.4"), variants.xhigh)
+    const upstream = await callThroughGateway("openai/gpt-5.4", opts)
+    expect(upstream?.reasoning_effort).toBe("xhigh")
+  })
+
+  test("legacy buggy key 'cloudflare-ai-gateway' does NOT reach the wire (proves the bug)", async () => {
+    // Sanity: confirms the bug class. If a future change accidentally restores
+    // providerID-keyed providerOptions, this test fails before users notice.
+    const upstream = await callThroughGateway("openai/gpt-5.4", {
+      "cloudflare-ai-gateway": { reasoningEffort: "high" },
+    })
+    expect(upstream?.reasoning_effort).toBeUndefined()
+  })
+})

packages/opencode/test/provider/transform.test.ts

@@ -2883,6 +2883,36 @@ describe("ProviderTransform.variants", () => {
       const result = ProviderTransform.variants(model)
       expect(Object.keys(result)).toEqual(["none", "minimal", "low", "medium", "high", "xhigh"])
     })
+
+    test("dotted gpt-5.x ids include 'minimal' (regression: matcher used to miss gpt-5.4)", () => {
+      const model = createMockModel({
+        id: "gpt-5.4",
+        providerID: "openai",
+        api: {
+          id: "gpt-5.4",
+          url: "https://api.openai.com",
+          npm: "@ai-sdk/openai",
+        },
+        release_date: "2026-03-05",
+      })
+      const result = ProviderTransform.variants(model)
+      expect(Object.keys(result)).toEqual(["none", "minimal", "low", "medium", "high", "xhigh"])
+    })
+
+    test("gpt-50 (lookalike) does not get gpt-5 family treatment", () => {
+      const model = createMockModel({
+        id: "gpt-50",
+        providerID: "openai",
+        api: {
+          id: "gpt-50",
+          url: "https://api.openai.com",
+          npm: "@ai-sdk/openai",
+        },
+        release_date: "2024-01-01",
+      })
+      const result = ProviderTransform.variants(model)
+      expect(Object.keys(result)).toEqual(["low", "medium", "high"])
+    })
   })
 
   describe("@ai-sdk/anthropic", () => {
@@ -3330,4 +3360,83 @@ describe("ProviderTransform.variants", () => {
       expect(result).toEqual({})
     })
   })
+
+  describe("ai-gateway-provider (cloudflare-ai-gateway)", () => {
+    const cfModel = (apiId: string, releaseDate = "2024-01-01") =>
+      createMockModel({
+        id: `cloudflare-ai-gateway/${apiId}`,
+        providerID: "cloudflare-ai-gateway",
+        api: {
+          id: apiId,
+          url: "https://gateway.ai.cloudflare.com/v1/compat",
+          npm: "ai-gateway-provider",
+        },
+        release_date: releaseDate,
+      })
+
+    test("openai gpt-5.4 includes xhigh effort (regression: variant=xhigh used to be silently ignored)", () => {
+      const result = ProviderTransform.variants(cfModel("openai/gpt-5.4", "2026-03-05"))
+      expect(result.xhigh).toEqual({ reasoningEffort: "xhigh" })
+      expect(result.high).toEqual({ reasoningEffort: "high" })
+      expect(Object.keys(result)).toContain("minimal")
+    })
+
+    test("openai gpt-5.2-codex includes xhigh", () => {
+      const result = ProviderTransform.variants(cfModel("openai/gpt-5.2-codex", "2025-12-11"))
+      expect(result.xhigh).toEqual({ reasoningEffort: "xhigh" })
+      expect(Object.keys(result)).toEqual(["low", "medium", "high", "xhigh"])
+    })
+
+    test("openai gpt-4o (no reasoning) returns empty", () => {
+      const model = cfModel("openai/gpt-4o")
+      model.capabilities.reasoning = false
+      const result = ProviderTransform.variants(model)
+      expect(result).toEqual({})
+    })
+
+    test("non-openai upstream falls back to widely-supported OAI efforts", () => {
+      const result = ProviderTransform.variants(cfModel("anthropic/claude-sonnet-4-6"))
+      expect(result).toEqual({
+        low: { reasoningEffort: "low" },
+        medium: { reasoningEffort: "medium" },
+        high: { reasoningEffort: "high" },
+      })
+    })
+  })
+})
+
+describe("ProviderTransform.providerOptions - ai-gateway-provider", () => {
+  const createModel = (overrides: Partial<any> = {}) =>
+    ({
+      id: "cloudflare-ai-gateway/openai/gpt-5.4",
+      providerID: "cloudflare-ai-gateway",
+      api: {
+        id: "openai/gpt-5.4",
+        url: "https://gateway.ai.cloudflare.com/v1/compat",
+        npm: "ai-gateway-provider",
+      },
+      capabilities: {
+        temperature: false,
+        reasoning: true,
+        attachment: true,
+        toolcall: true,
+        input: { text: true, audio: false, image: true, video: false, pdf: true },
+        output: { text: true, audio: false, image: false, video: false, pdf: false },
+        interleaved: false,
+      },
+      cost: { input: 1, output: 1, cache: { read: 0, write: 0 } },
+      limit: { context: 1_000_000, output: 128_000 },
+      status: "active",
+      options: {},
+      headers: {},
+      release_date: "2026-03-05",
+      ...overrides,
+    }) as any
+
+  test("routes options under openaiCompatible (the key @ai-sdk/openai-compatible reads)", () => {
+    // Regression: previously fell back to providerID="cloudflare-ai-gateway",
+    // which @ai-sdk/openai-compatible never reads, silently dropping reasoningEffort.
+    const result = ProviderTransform.providerOptions(createModel(), { reasoningEffort: "high" })
+    expect(result).toEqual({ openaiCompatible: { reasoningEffort: "high" } })
+  })
 })