feat(llm): cache-policy auto-placement (#26786)

packages/llm/test/provider/anthropic-messages-cache.recorded.test.ts

@@ -28,7 +28,12 @@ const recorded = recordedTests({
   provider: "anthropic",
   protocol: "anthropic-messages",
   requires: ["ANTHROPIC_API_KEY"],
-  options: { redactor: Redactor.defaults({ requestHeaders: { allow: ["content-type", "anthropic-version"] } }) },
+  // Two identical requests in one cassette — match by recording order so the
+  // second call replays the cached-hit interaction.
+  options: {
+    dispatch: "sequential",
+    redactor: Redactor.defaults({ requestHeaders: { allow: ["content-type", "anthropic-version"] } }),
+  },
 })
 
 describe("Anthropic Messages cache recorded", () => {

packages/llm/test/provider/bedrock-converse-cache.recorded.test.ts

@@ -35,6 +35,9 @@ const recorded = recordedTests({
   provider: "amazon-bedrock",
   protocol: "bedrock-converse",
   requires: ["AWS_ACCESS_KEY_ID", "AWS_SECRET_ACCESS_KEY"],
+  // Two identical requests in one cassette — match by recording order so the
+  // second call replays the cached-hit interaction.
+  options: { dispatch: "sequential" },
 })
 
 describe("Bedrock Converse cache recorded", () => {

packages/llm/test/provider/gemini-cache.recorded.test.ts

@@ -8,7 +8,7 @@ import { recordedTests } from "../recorded-test"
 
 const model = Gemini.model({
   id: "gemini-2.5-flash",
-  apiKey: process.env.GEMINI_API_KEY ?? "fixture",
+  apiKey: process.env.GOOGLE_GENERATIVE_AI_API_KEY ?? process.env.GEMINI_API_KEY ?? "fixture",
 })
 
 // Gemini does implicit prefix caching on 2.5+ models above ~1024 tokens. The
@@ -28,7 +28,10 @@ const recorded = recordedTests({
   prefix: "gemini-cache",
   provider: "google",
   protocol: "gemini",
-  requires: ["GEMINI_API_KEY"],
+  requires: ["GOOGLE_GENERATIVE_AI_API_KEY"],
+  // Two identical requests in one cassette — match by recording order so the
+  // second call replays the cached-hit interaction.
+  options: { dispatch: "sequential" },
 })
 
 describe("Gemini cache recorded", () => {

packages/llm/test/provider/openai-responses-cache.recorded.test.ts

@@ -29,6 +29,9 @@ const recorded = recordedTests({
   provider: "openai",
   protocol: "openai-responses",
   requires: ["OPENAI_API_KEY"],
+  // Two identical requests in one cassette — match by recording order so the
+  // second call replays the cached-hit interaction, not the cold-miss one.
+  options: { dispatch: "sequential" },
 })
 
 describe("OpenAI Responses cache recorded", () => {