fix(ai): validate cache tail counts

packages/ai/src/schema/options.ts

@@ -267,7 +267,7 @@ export const CachePolicyObject = Schema.Struct({
     Schema.Union([
       Schema.Literal("latest-user-message"),
       Schema.Literal("latest-assistant"),
-      Schema.Struct({ tail: Schema.Number }),
+      Schema.Struct({ tail: Schema.Natural }),
     ]),
   ),
   ttlSeconds: Schema.optional(Schema.Number),

packages/ai/test/cache-policy.test.ts

@@ -1,6 +1,6 @@
 import { describe, expect, test } from "bun:test"
-import { Effect } from "effect"
-import { CacheHint, LLM, Message } from "../src/index.js"
+import { Effect, Schema } from "effect"
+import { CacheHint, CachePolicyObject, LLM, Message } from "../src/index.js"
 import { Auth } from "../src/route.js"
 import { compileRequest } from "../src/route/client.js"
 import { AmazonBedrock, GoogleVertexMessages } from "../src/providers.js"
@@ -29,6 +29,37 @@ const geminiModel = Gemini.route
   })
   .model({ id: "gemini-2.5-flash" })
 
+const decodeCachePolicyObject = Schema.decodeUnknownSync(CachePolicyObject)
+
+describe("cache policy schema", () => {
+  test.each([0, 2])("accepts messages.tail count %d when decoding and constructing", (tail) => {
+    expect(decodeCachePolicyObject({ messages: { tail } })).toEqual({ messages: { tail } })
+    expect(
+      LLM.request({
+        model: anthropicModel,
+        prompt: "hi",
+        cache: { messages: { tail } },
+      }).cache,
+    ).toEqual({ messages: { tail } })
+  })
+
+  test.each([
+    ["negative", -1],
+    ["fraction", 1.5],
+    ["NaN", Number.NaN],
+    ["Infinity", Number.POSITIVE_INFINITY],
+  ])("rejects a %s messages.tail when decoding and constructing", (_name, tail) => {
+    expect(() => decodeCachePolicyObject({ messages: { tail } })).toThrow()
+    expect(() =>
+      LLM.request({
+        model: anthropicModel,
+        prompt: "hi",
+        cache: { messages: { tail } },
+      }),
+    ).toThrow()
+  })
+})
+
 describe("applyCachePolicy", () => {
   it.effect("undefined cache resolves to 'auto' (the recommended default)", () =>
     Effect.gen(function* () {
@@ -316,6 +347,19 @@ describe("applyCachePolicy", () => {
     }),
   )
 
+  test("messages: { tail: 0 } marks no message boundaries", () => {
+    const request = LLM.request({
+      model: anthropicModel,
+      messages: [Message.user("u1"), Message.assistant("a1")],
+      cache: { messages: { tail: 0 } },
+    })
+
+    expect(applyCachePolicy(request)).toBe(request)
+    expect(
+      request.messages.flatMap((message) => message.content.map((part) => ("cache" in part ? part.cache : undefined))),
+    ).toEqual([undefined, undefined])
+  })
+
   it.effect("'latest-assistant' marks the last assistant message", () =>
     Effect.gen(function* () {
       const prepared = yield* compileRequest(