fix(ai): emit cache markers for Alibaba Chat

packages/ai/src/cache-policy.ts

@@ -20,6 +20,7 @@ const AUTO: CachePolicyObject = {
   messages: { tail: 1 },
 }
 
+const MESSAGE_PREFIX: CachePolicyObject = { system: true, messages: { tail: 1 } }
 const NONE: CachePolicyObject = {}
 const BREAKPOINT_CAP = 4
 
@@ -28,8 +29,8 @@ const BREAKPOINT_CAP = 4
 //   - "auto"      → tools + first/last system + final message boundary.
 //   - "none"      → no auto placement; manual `CacheHint`s still flow.
 //   - object form → exactly what the caller asked for.
-const resolve = (policy: CachePolicy | undefined): CachePolicyObject => {
-  if (policy === undefined || policy === "auto") return AUTO
+const resolve = (policy: CachePolicy | undefined, automatic = AUTO): CachePolicyObject => {
+  if (policy === undefined || policy === "auto") return automatic
   if (policy === "none") return NONE
   return policy
 }
@@ -38,6 +39,7 @@ const resolve = (policy: CachePolicy | undefined): CachePolicyObject => {
 // prefix caching, Gemini's implicit + out-of-band CachedContent). Skip the
 // whole policy pass for these — emitting hints would be harmless but pointless.
 const RESPECTS_INLINE_HINTS = new Set([
+  "alibaba-chat",
   "alibaba-messages",
   "anthropic-messages",
   "anthropic-compatible-messages",
@@ -58,7 +60,7 @@ const openRouterPolicy = (modelID: string): CachePolicyObject => {
   // `~anthropic/claude-sonnet-latest` style IDs are OpenRouter aliases for the latest model in a family.
   const id = modelID.replace(/^~/, "")
   if (id.startsWith("anthropic/")) return AUTO
-  if (id.startsWith("qwen/")) return { system: true, messages: { tail: 1 } }
+  if (id.startsWith("qwen/")) return MESSAGE_PREFIX
   return NONE
 }
 
@@ -152,8 +154,8 @@ const markMessages = (
   return next
 }
 
-const countHints = (request: LLMRequest) =>
-  countToolHints(request.tools) +
+const countHints = (request: LLMRequest, tools = true) =>
+  (tools ? countToolHints(request.tools) : 0) +
   request.system.reduce((count, part) => count + (part.cache === undefined ? 0 : 1), 0) +
   request.messages.reduce(
     (count, message) =>
@@ -167,15 +169,16 @@ const countHints = (request: LLMRequest) =>
 
 export const applyCachePolicy = (request: LLMRequest): LLMRequest => {
   if (!RESPECTS_INLINE_HINTS.has(request.model.route.id)) return request
-  const policy =
-    request.model.route.id === "openrouter" && (request.cache === undefined || request.cache === "auto")
-      ? openRouterPolicy(request.model.id)
-      : resolve(request.cache)
+  const cacheTools = request.model.route.id !== "alibaba-chat"
+  const policy = resolve(
+    request.cache,
+    request.model.route.id === "openrouter" ? openRouterPolicy(request.model.id) : cacheTools ? AUTO : MESSAGE_PREFIX,
+  )
   if (!policy.tools && !policy.system && !policy.messages) return request
 
   const hint = makeHint(policy.ttlSeconds)
-  const budget = { remaining: Math.max(0, BREAKPOINT_CAP - countHints(request)) }
-  const tools = policy.tools ? markLastTool(request.tools, hint, budget) : request.tools
+  const budget = { remaining: Math.max(0, BREAKPOINT_CAP - countHints(request, cacheTools)) }
+  const tools = policy.tools && cacheTools ? markLastTool(request.tools, hint, budget) : request.tools
   const system = policy.system ? markSystemBoundaries(request.system, hint, budget) : request.system
   const messages = policy.messages ? markMessages(request.messages, policy.messages, hint, budget) : request.messages
 

packages/ai/src/protocols/alibaba-chat.ts

@@ -4,6 +4,7 @@ import type { LanguageModelCompatibility } from "../schema/index.js"
 import { OpenAIChat } from "./openai-chat.js"
 import { JsonObject, ProviderShared } from "./shared.js"
 import { OpenResponsesOptions } from "./utils/open-responses-options.js"
+import { newBreakpoints } from "./utils/cache.js"
 
 export type ReasoningEffort = OpenResponsesOptions.ReasoningEffort
 
@@ -67,8 +68,17 @@ export const protocol = Protocol.make({
     }),
     from: Effect.fn("AlibabaChat.fromRequest")(function* (req) {
       const opts = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Options))(req.providerOptions ?? {})
+      const breakpoints = newBreakpoints(4)
       return {
-        ...(yield* OpenAIChat.protocol.body.from(req)),
+        ...(yield* OpenAIChat.fromRequest(req, {
+          // Alibaba caches tool definitions with system content and only supports 5-minute markers.
+          cacheTools: false,
+          cacheControl: (cache) => {
+            if (cache === undefined || breakpoints.remaining === 0) return undefined
+            breakpoints.remaining -= 1
+            return { type: "ephemeral" }
+          },
+        })),
         enable_thinking: opts.enableThinking,
         // Alibaba also rejects an explicit budget that is not below `max_completion_tokens`.
         thinking_budget:

packages/ai/src/protocols/openai-chat.ts

@@ -330,6 +330,7 @@ export interface ParserState {
 // fields into `LLMRequest`.
 interface LoweringOptions {
   readonly cacheControl?: (cache: CacheHint | undefined) => OpenAIChatCacheControl | undefined
+  readonly cacheTools?: boolean
   readonly toolCallID?: (id: string) => string
 }
 
@@ -341,7 +342,7 @@ const lowerTool = (tool: ToolDefinition, options: LoweringOptions, supportsStric
     parameters: tool.inputSchema,
     ...(supportsStrictMode ? { strict: false } : {}),
   },
-  cache_control: options.cacheControl?.(tool.cache),
+  cache_control: options.cacheTools === false ? undefined : options.cacheControl?.(tool.cache),
 })
 
 const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>

packages/ai/test/provider/alibaba-capabilities.recorded.test.ts

@@ -17,6 +17,8 @@ const record = (api: "chat" | "messages" | "responses") =>
   })
 
 for (const api of ["chat", "messages", "responses"] as const) {
+  // Preserve the request shape of Chat recordings made before explicit caching was enabled.
+  const cache = api === "chat" ? "none" : undefined
   const recorded = record(api)
   describe(`Alibaba ${api} capabilities`, () => {
     for (const enabled of [false, true]) {
@@ -27,6 +29,7 @@ for (const api of ["chat", "messages", "responses"] as const) {
           Effect.gen(function* () {
             const request = LLM.request({
               model: alibaba[api]("qwen3.7-plus"),
+              cache,
               providerOptions:
                 api === "messages"
                   ? { thinking: { type: enabled ? "enabled" : "disabled", ...(enabled ? { budgetTokens: 1024 } : {}) } }
@@ -61,6 +64,7 @@ for (const api of ["chat", "messages", "responses"] as const) {
           const response = yield* LLMClient.generate(
             LLM.request({
               model: alibaba[api]("qwen3.8-flash"),
+              cache,
               providerOptions: api === "messages" ? { thinking: { type: "disabled" } } : { enableThinking: false },
               messages: [
                 Message.user([
@@ -84,6 +88,7 @@ for (const api of ["chat", "messages", "responses"] as const) {
           const response = yield* LLMClient.generate(
             LLM.request({
               model: alibaba[api]("qwen3.8-max"),
+              cache,
               prompt: "Find the current weather in Paris.",
               providerOptions: api === "messages" ? { thinking: { type: "disabled" } } : { reasoningEffort: "none" },
               tools: [
@@ -118,6 +123,7 @@ record("chat").effect.with(
       const response = yield* LLMClient.generate(
         LLM.request({
           model: alibaba.chat("qwen3.8-max"),
+          cache: "none",
           prompt: 'Return a JSON object with one key "city" set to the capital city of France.',
           providerOptions: { reasoningEffort: "none", responseFormat: { type: "json_object" } },
           generation: { maxTokens: 1024 },

packages/ai/test/provider/alibaba.recorded.test.ts

@@ -19,6 +19,8 @@ const weather = ToolDefinition.make({
 })
 
 for (const api of ["chat", "messages", "responses"] as const) {
+  // Preserve the request shape of Chat recordings made before explicit caching was enabled.
+  const cache = api === "chat" ? "none" : undefined
   const recorded = recordedTests({
     prefix: `alibaba-${api}`,
     provider: "alibaba",
@@ -37,6 +39,7 @@ for (const api of ["chat", "messages", "responses"] as const) {
           Effect.gen(function* () {
             const request = LLM.request({
               model: alibaba[api]("qwen3.8-max"),
+              cache,
               prompt: "What is 173 multiplied by 219? Reply with only the final integer.",
               providerOptions: api === "messages" ? { effort } : { reasoningEffort: effort },
               generation: { maxTokens: 4096 },
@@ -68,6 +71,7 @@ for (const api of ["chat", "messages", "responses"] as const) {
         Effect.gen(function* () {
           const request = LLM.request({
             model: alibaba[api]("qwen3.8-max"),
+            cache,
             providerOptions:
               api === "messages"
                 ? { effort: "medium" }