From 3c4940b5f1138d435557742600434f220f359eac Mon Sep 17 00:00:00 2001
From: peiwenz2 <110538780+peiwenz2@users.noreply.github.com>
Date: Sun, 24 May 2026 20:36:14 +0800
Subject: [PATCH 1/2] fix(provider): enable prompt caching for DashScope-routed
 Qwen models
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The applyCaching gate in provider/transform.ts only fired for @ai-sdk/anthropic
and @ai-sdk/alibaba. opencode's catalog wires every DashScope (Alibaba Cloud
Model Studio / Bailian) model — alibaba, alibaba-cn, alibaba-coding-plan(-cn)
— through @ai-sdk/openai-compatible pointing at
https://dashscope[-intl].aliyuncs.com/compatible-mode/v1, so no cache_control
markers were ever sent on the wire.

Without caching, browser tasks on qwen3-max pay full price every turn and
the model can end up more expensive than Opus 4.7 — exactly the cost cliff
hinted at in the model recommendations. Bailian's cache_control protocol is
shaped like Anthropic's (5m TTL, 4-breakpoint cap, 10% cache_read /
125% cache_write), so the existing 4-marker strategy carries over without
needing a separate code path.

Changes:
- Add isDashScopeRoutedModel() helper covering alibaba*, dashscope, bailian
  and the @ai-sdk/alibaba SDK path. Include it in the applyCaching gate.
- DashScope wants cache_control on a content block, not on the message
  envelope. System messages arrive here as strings; lift them into a
  single-block array before the content-level marker is applied so the
  AI SDK openai-compatible plugin can spread the cache_control field onto
  the wire block (per
  packages/openai-compatible/src/chat/convert-to-openai-compatible-chat-messages.ts
  in vercel/ai).
- Add @ai-sdk/alibaba to sdkKey() so persisted providerOptions stored under
  the catalog's providerID (e.g. alibaba-cn) remap to the SDK's "alibaba"
  namespace on session reload.
- Regression tests in test/provider/dashscope-cache.test.ts cover the
  alibaba-cn / alibaba / alibaba-coding-plan-cn paths, the 4-breakpoint
  cap, the string-to-block lift, the @ai-sdk/alibaba SDK path, and the
  gateway-exclusion contract.

Refs: https://www.alibabacloud.com/help/zh/model-studio/context-cache
---
 packages/opencode/src/provider/transform.ts   |  56 +++-
 .../test/provider/dashscope-cache.test.ts     | 300 ++++++++++++++++++
 2 files changed, 355 insertions(+), 1 deletion(-)
 create mode 100644 packages/opencode/test/provider/dashscope-cache.test.ts

diff --git a/packages/opencode/src/provider/transform.ts b/packages/opencode/src/provider/transform.ts
index 56a35d9af..85d0cbe2c 100644
--- a/packages/opencode/src/provider/transform.ts
+++ b/packages/opencode/src/provider/transform.ts
@@ -44,6 +44,11 @@ function sdkKey(npm: string): string | undefined {
       return "gateway"
     case "@openrouter/ai-sdk-provider":
       return "openrouter"
+    case "@ai-sdk/alibaba":
+      // Persisted messages store providerOptions under the stored providerID
+      // (e.g. "alibaba-cn"); remap to the SDK-expected "alibaba" namespace so
+      // cacheControl set in `applyCaching` survives a session reload.
+      return "alibaba"
     case "ai-gateway-provider":
       // ai-gateway-provider/unified wraps createOpenAICompatible({ name: "Unified" }),
       // and @ai-sdk/openai-compatible parses compatibleOptions from one of
@@ -338,6 +343,27 @@ function normalizeMessages(
   return msgs
 }
 
+// Detect models routed through DashScope (Alibaba Cloud Model Studio /
+// Bailian / Tongyi) regardless of which AI SDK package the catalog mapped
+// them to. opencode's snapshot today ships `alibaba`, `alibaba-cn`,
+// `alibaba-coding-plan(-cn)` etc. as `@ai-sdk/openai-compatible` pointing at
+// `https://dashscope[-intl].aliyuncs.com/compatible-mode/v1`, but the
+// upstream cache-control protocol is the same one `@ai-sdk/alibaba` speaks:
+// https://www.alibabacloud.com/help/zh/model-studio/context-cache .
+//
+// Without this check, `applyCaching` is skipped for DashScope users — every
+// turn re-sends the whole conversation at full price and the model looks far
+// more expensive than it actually is. (Bailian bills cached input at 10% of
+// the standard rate, cache writes at 125%, with a 5-minute TTL and a hard
+// cap of 4 markers per request — the same shape `applyCaching` already emits
+// for Anthropic, so the existing strategy carries over.)
+function isDashScopeRoutedModel(model: Provider.Model): boolean {
+  if (model.providerID.startsWith("alibaba")) return true
+  if (model.providerID === "dashscope" || model.providerID === "bailian") return true
+  if (model.api.npm === "@ai-sdk/alibaba") return true
+  return false
+}
+
 function applyCaching(msgs: ModelMessage[], model: Provider.Model): ModelMessage[] {
   const system = msgs.filter((msg) => msg.role === "system").slice(0, 2)
   const final = msgs.filter((msg) => msg.role !== "system").slice(-2)
@@ -363,7 +389,27 @@ function applyCaching(msgs: ModelMessage[], model: Provider.Model): ModelMessage
     },
   }
 
+  // DashScope requires `cache_control` on a content *block*, not on the
+  // message envelope. `@ai-sdk/openai-compatible` lifts each part's
+  // `providerOptions.openaiCompatible.*` onto the corresponding wire block
+  // (see vercel/ai `convert-to-openai-compatible-chat-messages.ts`), so the
+  // existing content-level path produces the exact wire shape Bailian
+  // documents — *if* the message has array content to attach a part to.
+  // System messages still arrive here as strings (see `session/llm.ts`
+  // line 152), so lift them into a single-block array first.
+  const requiresBlockLevel = isDashScopeRoutedModel(model)
+
   for (const msg of unique([...system, ...final])) {
+    if (requiresBlockLevel && typeof msg.content === "string" && msg.content.length > 0) {
+      // The AI SDK v3 ModelMessage union types system content as `string`, so
+      // assigning an array tripped tsc. The runtime accepts array content for
+      // every role we touch here (verified against `@ai-sdk/openai-compatible`'s
+      // `convert-to-openai-compatible-chat-messages.ts`, which passes
+      // `content` through verbatim for system and maps each part for
+      // user/assistant). Cast through `any` for the assignment only.
+      ;(msg as any).content = [{ type: "text", text: msg.content }]
+    }
+
     const useMessageLevelOptions =
       model.providerID === "anthropic" ||
       model.providerID.includes("bedrock") ||
@@ -438,7 +484,15 @@ export function message(msgs: ModelMessage[], model: Provider.Model, options: Re
       model.id.includes("anthropic") ||
       model.id.includes("claude") ||
       model.api.npm === "@ai-sdk/anthropic" ||
-      model.api.npm === "@ai-sdk/alibaba") &&
+      model.api.npm === "@ai-sdk/alibaba" ||
+      // DashScope-routed models (qwen3-max, qwen3.6-*, qwen3-coder-plus,
+      // kimi-k2.5 on alibaba-cn, deepseek-v3.2 on alibaba-cn, …) all speak
+      // the same cache_control protocol as the native `@ai-sdk/alibaba`
+      // SDK even though the catalog wires them through
+      // `@ai-sdk/openai-compatible`. Without this branch, browser tasks on
+      // Qwen pay full price every turn — which is exactly the scenario
+      // README describes as "more expensive than Opus 4.7 without caching".
+      isDashScopeRoutedModel(model)) &&
     model.api.npm !== "@ai-sdk/gateway"
   ) {
     msgs = applyCaching(msgs, model)
diff --git a/packages/opencode/test/provider/dashscope-cache.test.ts b/packages/opencode/test/provider/dashscope-cache.test.ts
new file mode 100644
index 000000000..22fd04c58
--- /dev/null
+++ b/packages/opencode/test/provider/dashscope-cache.test.ts
@@ -0,0 +1,300 @@
+import { describe, expect, test } from "bun:test"
+import { ProviderTransform } from "@/provider/transform"
+
+// Regression coverage for the DashScope (Alibaba Cloud Model Studio / Bailian
+// / Tongyi) prompt-cache path.
+//
+// The catalog wires `alibaba`, `alibaba-cn`, `alibaba-coding-plan(-cn)`
+// through `@ai-sdk/openai-compatible` pointing at
+// `https://dashscope[-intl].aliyuncs.com/compatible-mode/v1`. Prior to this
+// patch, `ProviderTransform.message` skipped `applyCaching` for those models
+// — even though the upstream service speaks the same cache_control protocol
+// `@ai-sdk/alibaba` does
+// (https://www.alibabacloud.com/help/zh/model-studio/context-cache). Browser
+// agent loops therefore re-sent the full conversation every turn at full
+// price, making qwen3-max look more expensive than Opus 4.7.
+//
+// These tests pin three properties:
+//   1. The gate fires for every DashScope-routed providerID.
+//   2. `cache_control` lands on a content *block*, not on the message
+//      envelope — Bailian rejects it at message level.
+//   3. System messages, which arrive as plain strings from `session/llm.ts`,
+//      get lifted into a single-block array so the block-level path applies.
+
+type AnyModel = Parameters<typeof ProviderTransform.message>[1]
+
+const dashscopeOpenAICompatibleModel = (overrides: Partial<any> = {}): AnyModel =>
+  ({
+    id: "alibaba-cn/qwen3-max",
+    providerID: "alibaba-cn",
+    api: {
+      id: "qwen3-max",
+      url: "https://dashscope.aliyuncs.com/compatible-mode/v1",
+      npm: "@ai-sdk/openai-compatible",
+    },
+    name: "Qwen3 Max",
+    capabilities: {
+      temperature: true,
+      reasoning: false,
+      attachment: false,
+      toolcall: true,
+      input: { text: true, audio: false, image: false, video: false, pdf: false },
+      output: { text: true, audio: false, image: false, video: false, pdf: false },
+      interleaved: false,
+    },
+    cost: { input: 1.2, output: 6 },
+    limit: { context: 262144, output: 65536 },
+    status: "active",
+    options: {},
+    headers: {},
+    ...overrides,
+  }) as any
+
+describe("ProviderTransform.message - DashScope (alibaba-*) cache_control", () => {
+  test("alibaba-cn via openai-compatible: cache_control lands on the last content block of system+last messages", () => {
+    const model = dashscopeOpenAICompatibleModel()
+    const msgs = [
+      { role: "system", content: "You are a helpful browser agent." },
+      { role: "system", content: "Tools: browser_execute, search." },
+      { role: "user", content: "Find flights from NYC to SF tomorrow." },
+      { role: "assistant", content: [{ type: "text", text: "I'll start by searching." }] },
+    ] as any[]
+
+    const result = ProviderTransform.message(msgs, model, {}) as any[]
+
+    // System messages were lifted from string → array so cache_control can
+    // ride on a content block.
+    expect(Array.isArray(result[0].content)).toBe(true)
+    expect(Array.isArray(result[1].content)).toBe(true)
+
+    // The wire-relevant marker for DashScope is `openaiCompatible.cache_control`
+    // (snake_case) — the SDK spreads this onto the content block; the other
+    // namespaces stay attached so the same loop output can switch providers
+    // without remarking. Allow extra namespaces (anthropic/openrouter/etc.)
+    // since they are documented dead weight when the wire goes via OAI-compat.
+    for (const i of [0, 1, 2, 3]) {
+      const lastBlock = result[i].content[result[i].content.length - 1]
+      expect(lastBlock.providerOptions.openaiCompatible.cache_control).toEqual({
+        type: "ephemeral",
+      })
+      // The Alibaba SDK namespace must also be set for the case where a user
+      // pins `@ai-sdk/alibaba` instead of openai-compatible — same one-pass
+      // marking, no second transform needed.
+      expect(lastBlock.providerOptions.alibaba.cacheControl).toEqual({
+        type: "ephemeral",
+      })
+      // Cache_control must NOT leak onto the message envelope; Bailian will
+      // not look there and would still create a fresh prefix every turn.
+      expect(result[i].providerOptions?.openaiCompatible?.cache_control).toBeUndefined()
+    }
+  })
+
+  test("alibaba (intl) is treated the same as alibaba-cn", () => {
+    const model = dashscopeOpenAICompatibleModel({
+      providerID: "alibaba",
+      api: {
+        id: "qwen3-max",
+        url: "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
+        npm: "@ai-sdk/openai-compatible",
+      },
+    })
+    const msgs = [
+      { role: "system", content: "system prompt" },
+      { role: "user", content: "user message" },
+    ] as any[]
+
+    const result = ProviderTransform.message(msgs, model, {}) as any[]
+
+    expect(Array.isArray(result[0].content)).toBe(true)
+    expect(result[0].content[0].providerOptions.openaiCompatible.cache_control).toEqual({
+      type: "ephemeral",
+    })
+  })
+
+  test("alibaba-coding-plan-cn (subscription tier) also gets cache_control", () => {
+    const model = dashscopeOpenAICompatibleModel({
+      providerID: "alibaba-coding-plan-cn",
+      id: "alibaba-coding-plan-cn/qwen3-coder-plus",
+      api: {
+        id: "qwen3-coder-plus",
+        url: "https://dashscope.aliyuncs.com/compatible-mode/v1",
+        npm: "@ai-sdk/openai-compatible",
+      },
+    })
+    const msgs = [
+      { role: "system", content: "system prompt" },
+      { role: "user", content: "user message" },
+    ] as any[]
+
+    const result = ProviderTransform.message(msgs, model, {}) as any[]
+
+    const block = result[0].content[0]
+    expect(block.providerOptions.openaiCompatible.cache_control).toEqual({ type: "ephemeral" })
+  })
+
+  test("at most 4 markers — first 2 system + last 2 non-system (matches Bailian's 4-breakpoint cap)", () => {
+    const model = dashscopeOpenAICompatibleModel()
+    const msgs = [
+      { role: "system", content: "sys A (marked)" },
+      { role: "system", content: "sys B (marked)" },
+      { role: "system", content: "sys C (NOT marked)" }, // beyond first 2
+      { role: "user", content: "u1 (NOT marked)" },
+      { role: "assistant", content: [{ type: "text", text: "a1 (NOT marked)" }] },
+      { role: "user", content: "u2 (marked - second to last)" },
+      { role: "assistant", content: [{ type: "text", text: "a2 (marked - last)" }] },
+    ] as any[]
+
+    const result = ProviderTransform.message(msgs, model, {}) as any[]
+
+    const hasMarker = (m: any) => {
+      if (Array.isArray(m.content) && m.content.length > 0) {
+        const last = m.content[m.content.length - 1]
+        return last?.providerOptions?.openaiCompatible?.cache_control != null
+      }
+      return m.providerOptions?.openaiCompatible?.cache_control != null
+    }
+
+    expect(hasMarker(result[0])).toBe(true) // sys A
+    expect(hasMarker(result[1])).toBe(true) // sys B
+    expect(hasMarker(result[2])).toBe(false) // sys C — beyond the first 2 system slots
+    expect(hasMarker(result[3])).toBe(false) // u1 — not in the last 2
+    expect(hasMarker(result[4])).toBe(false) // a1 — not in the last 2
+    expect(hasMarker(result[5])).toBe(true) // u2 — second-to-last non-system
+    expect(hasMarker(result[6])).toBe(true) // a2 — last non-system
+  })
+
+  test("string system content survives sanitization and gets a real text block", () => {
+    const model = dashscopeOpenAICompatibleModel()
+    const msgs = [{ role: "system", content: "hello world" }, { role: "user", content: "hi" }] as any[]
+
+    const result = ProviderTransform.message(msgs, model, {}) as any[]
+
+    expect(result[0].content).toEqual([
+      {
+        type: "text",
+        text: "hello world",
+        providerOptions: expect.objectContaining({
+          openaiCompatible: { cache_control: { type: "ephemeral" } },
+          alibaba: { cacheControl: { type: "ephemeral" } },
+        }),
+      },
+    ])
+  })
+
+  test("empty system content is left alone (no spurious empty block, no marker)", () => {
+    const model = dashscopeOpenAICompatibleModel()
+    // sanitizeSurrogates collapses to "" then the system stays as empty
+    // string; we should not generate a `[{ type: "text", text: "" }]` block.
+    const msgs = [{ role: "system", content: "" }, { role: "user", content: "hi" }] as any[]
+
+    const result = ProviderTransform.message(msgs, model, {}) as any[]
+
+    expect(result[0].content).toBe("")
+  })
+
+  test("@ai-sdk/alibaba SDK path still works (defense in depth)", () => {
+    const model = dashscopeOpenAICompatibleModel({
+      api: {
+        id: "qwen3-max",
+        url: "https://dashscope.aliyuncs.com/compatible-mode/v1",
+        npm: "@ai-sdk/alibaba",
+      },
+    })
+    const msgs = [
+      { role: "system", content: "sys" },
+      { role: "user", content: [{ type: "text", text: "u" }] },
+    ] as any[]
+
+    const result = ProviderTransform.message(msgs, model, {}) as any[]
+
+    expect(result[0].content[0].providerOptions.alibaba.cacheControl).toEqual({
+      type: "ephemeral",
+    })
+    expect(result[1].content[0].providerOptions.alibaba.cacheControl).toEqual({
+      type: "ephemeral",
+    })
+  })
+
+  test("gateway-routed alibaba is excluded (gateway handles caching itself)", () => {
+    const model = dashscopeOpenAICompatibleModel({
+      providerID: "vercel",
+      api: {
+        id: "alibaba/qwen3-max",
+        url: "https://ai-gateway.vercel.sh/v3/ai",
+        npm: "@ai-sdk/gateway",
+      },
+    })
+    const msgs = [
+      { role: "system", content: "sys" },
+      { role: "user", content: "u" },
+    ] as any[]
+
+    const result = ProviderTransform.message(msgs, model, {}) as any[]
+
+    // Gateway path must NOT inject cache_control — the gateway provider does
+    // its own caching via `gateway: { caching: "auto" }` set in
+    // ProviderTransform.options. Double-marking would risk double-billing.
+    expect(result[0].content).toBe("sys")
+    expect(result[0].providerOptions).toBeUndefined()
+  })
+
+  test("non-DashScope openai-compatible provider (e.g. deepseek direct) is still excluded", () => {
+    const model = dashscopeOpenAICompatibleModel({
+      providerID: "deepseek",
+      id: "deepseek/deepseek-chat",
+      api: {
+        id: "deepseek-chat",
+        url: "https://api.deepseek.com/v1",
+        npm: "@ai-sdk/openai-compatible",
+      },
+    })
+    const msgs = [
+      { role: "system", content: "sys" },
+      { role: "user", content: "u" },
+    ] as any[]
+
+    const result = ProviderTransform.message(msgs, model, {}) as any[]
+
+    // DeepSeek's direct API uses implicit caching — no per-message
+    // cache_control to set. Keep the current no-op behaviour for it; if
+    // someone routes deepseek through DashScope's alibaba-cn provider, that
+    // model's providerID will be "alibaba-cn" and the gate above fires.
+    expect(result[0].content).toBe("sys")
+    expect(result[0].providerOptions).toBeUndefined()
+  })
+})
+
+describe("ProviderTransform.message - providerOptions key remap for @ai-sdk/alibaba", () => {
+  test("stored providerID 'alibaba-cn' remaps to SDK key 'alibaba' when api.npm is @ai-sdk/alibaba", () => {
+    // Sessions that were originally written with providerID `alibaba-cn`
+    // persist providerOptions under that exact key. When the SDK switches
+    // to `@ai-sdk/alibaba`, sdkKey() must remap to "alibaba" so the
+    // serialized options reach the provider plugin.
+    const model = dashscopeOpenAICompatibleModel({
+      api: {
+        id: "qwen3-max",
+        url: "https://dashscope.aliyuncs.com/compatible-mode/v1",
+        npm: "@ai-sdk/alibaba",
+      },
+    })
+    const msgs = [
+      {
+        role: "user",
+        content: [
+          {
+            type: "text",
+            text: "hello",
+            providerOptions: {
+              "alibaba-cn": { cacheControl: { type: "ephemeral" } },
+            },
+          },
+        ],
+      },
+    ] as any[]
+
+    const result = ProviderTransform.message(msgs, model, {}) as any[]
+    const block = result[0].content[0]
+    expect(block.providerOptions.alibaba?.cacheControl).toEqual({ type: "ephemeral" })
+    expect(block.providerOptions["alibaba-cn"]).toBeUndefined()
+  })
+})

From b49bbefcf540840b27929d5860f067d3f990ed9b Mon Sep 17 00:00:00 2001
From: peiwenz2 <110538780+peiwenz2@users.noreply.github.com>
Date: Sun, 24 May 2026 22:25:39 +0800
Subject: [PATCH 2/2] test(provider): rename example model in cache regression
 tests to qwen3.7-max
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

qwen3.7-max is the model the README pricing footnote calls out — and the
one users have actually been hitting the uncached cost cliff on. The cache
gate added in the previous commit is keyed on providerID (alibaba* /
dashscope / bailian / @ai-sdk/alibaba), so it already covers every model
the catalog hangs under those providers, including qwen3.7-max, the
qwen3.7-max-2026-05-20 snapshot, qwen3-max, qwen3.6-*, kimi-k2.5, and
deepseek-v3.2.

This commit only renames the example model id in the fixture so the test
output makes that coverage visible at a glance to anyone reading the diff.
No production code changes; cost numbers in the fixture are informational —
the gate does not read cost. Live probe against
dashscope.aliyuncs.com/compatible-mode/v1 on 2026-05-24 confirmed
qwen3.7-max returns prompt_tokens_details.cache_creation /
cache_creation_input_tokens / cache_type=ephemeral / cached_tokens in
response usage, so the wire-format path the gate enables is honored
server-side.
---
 .../test/provider/dashscope-cache.test.ts     | 24 ++++++++++++-------
 1 file changed, 15 insertions(+), 9 deletions(-)

diff --git a/packages/opencode/test/provider/dashscope-cache.test.ts b/packages/opencode/test/provider/dashscope-cache.test.ts
index 22fd04c58..8b48a1925 100644
--- a/packages/opencode/test/provider/dashscope-cache.test.ts
+++ b/packages/opencode/test/provider/dashscope-cache.test.ts
@@ -12,7 +12,7 @@ import { ProviderTransform } from "@/provider/transform"
 // `@ai-sdk/alibaba` does
 // (https://www.alibabacloud.com/help/zh/model-studio/context-cache). Browser
 // agent loops therefore re-sent the full conversation every turn at full
-// price, making qwen3-max look more expensive than Opus 4.7.
+// price, making qwen3.7-max look more expensive than Opus 4.7.
 //
 // These tests pin three properties:
 //   1. The gate fires for every DashScope-routed providerID.
@@ -25,23 +25,29 @@ type AnyModel = Parameters<typeof ProviderTransform.message>[1]
 
 const dashscopeOpenAICompatibleModel = (overrides: Partial<any> = {}): AnyModel =>
   ({
-    id: "alibaba-cn/qwen3-max",
+    id: "alibaba-cn/qwen3.7-max",
     providerID: "alibaba-cn",
     api: {
-      id: "qwen3-max",
+      id: "qwen3.7-max",
       url: "https://dashscope.aliyuncs.com/compatible-mode/v1",
       npm: "@ai-sdk/openai-compatible",
     },
-    name: "Qwen3 Max",
+    name: "Qwen3.7 Max",
     capabilities: {
       temperature: true,
-      reasoning: false,
+      // qwen3.7-max returns reasoning_tokens by default per a live probe
+      // against dashscope.aliyuncs.com on 2026-05-24 — mirror the catalog
+      // convention for thinking models.
+      reasoning: true,
       attachment: false,
       toolcall: true,
       input: { text: true, audio: false, image: false, video: false, pdf: false },
       output: { text: true, audio: false, image: false, video: false, pdf: false },
       interleaved: false,
     },
+    // The cache gate is providerID-keyed and does not read cost, so these
+    // numbers are informational only — once qwen3.7-max lands in models.dev
+    // the snapshot will overwrite the catalog cost field anyway.
     cost: { input: 1.2, output: 6 },
     limit: { context: 262144, output: 65536 },
     status: "active",
@@ -93,7 +99,7 @@ describe("ProviderTransform.message - DashScope (alibaba-*) cache_control", () =
     const model = dashscopeOpenAICompatibleModel({
       providerID: "alibaba",
       api: {
-        id: "qwen3-max",
+        id: "qwen3.7-max",
         url: "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
         npm: "@ai-sdk/openai-compatible",
       },
@@ -195,7 +201,7 @@ describe("ProviderTransform.message - DashScope (alibaba-*) cache_control", () =
   test("@ai-sdk/alibaba SDK path still works (defense in depth)", () => {
     const model = dashscopeOpenAICompatibleModel({
       api: {
-        id: "qwen3-max",
+        id: "qwen3.7-max",
         url: "https://dashscope.aliyuncs.com/compatible-mode/v1",
         npm: "@ai-sdk/alibaba",
       },
@@ -219,7 +225,7 @@ describe("ProviderTransform.message - DashScope (alibaba-*) cache_control", () =
     const model = dashscopeOpenAICompatibleModel({
       providerID: "vercel",
       api: {
-        id: "alibaba/qwen3-max",
+        id: "alibaba/qwen3.7-max",
         url: "https://ai-gateway.vercel.sh/v3/ai",
         npm: "@ai-sdk/gateway",
       },
@@ -272,7 +278,7 @@ describe("ProviderTransform.message - providerOptions key remap for @ai-sdk/alib
     // serialized options reach the provider plugin.
     const model = dashscopeOpenAICompatibleModel({
       api: {
-        id: "qwen3-max",
+        id: "qwen3.7-max",
         url: "https://dashscope.aliyuncs.com/compatible-mode/v1",
         npm: "@ai-sdk/alibaba",
       },