fix(cli): gate prompt cache breakpoints to first-party OpenAI providers

.changeset/custom-provider-cache-breakpoint.md

@@ -0,0 +1,5 @@
+---
+"@kilocode/cli": patch
+---
+
+Stop sending prompt_cache_breakpoint to custom OpenAI-compatible providers, and to first-party provider IDs rerouted through custom endpoint overrides; both reject the parameter with HTTP 400.

packages/opencode/src/provider/transform.ts

@@ -367,8 +367,29 @@ function isLikelyChatGPTSubscription(model: Provider.Model): boolean {
   return model.providerID === "openai" && model.cost?.input === 0 && model.cost?.output === 0
 }
 
-function supportsPromptCacheBreakpoint(model: Provider.Model): boolean {
+// Endpoint overrides (options.endpoint / options.baseURL) reroute a first-party
+// provider ID through a proxy that may reject prompt_cache_breakpoint (#13285).
+function isFirstPartyBreakpointEndpoint(model: Provider.Model, options: Record<string, unknown>): boolean {
+  const override = options["providerEndpointOverride"] ?? options["endpoint"] ?? options["baseURL"]
+  if (typeof override !== "string" || override.length === 0) return true
+  let host: string
+  try {
+    host = new URL(override).hostname.toLowerCase()
+  } catch {
+    return false
+  }
+  if (model.providerID === "openai") return host === "openai.com" || host.endsWith(".openai.com")
+  if (model.providerID === "azure")
+    return [".azure.com", ".azure.us", ".azure.cn", ".azure-api.net"].some((s) => host.endsWith(s))
+  return false
+}
+
+function supportsPromptCacheBreakpoint(model: Provider.Model, options: Record<string, unknown> = {}): boolean {
   if (isLikelyChatGPTSubscription(model)) return false
+  // Only first-party OpenAI-family deployments support explicit breakpoints;
+  // custom @ai-sdk/openai endpoints reject prompt_cache_breakpoint (#13285).
+  if (!["openai", "azure", "kilo"].includes(model.providerID)) return false
+  if (!isFirstPartyBreakpointEndpoint(model, options)) return false
   const match = model.api.id.match(/gpt-(\d+)\.(\d+)/)
   if (match) {
     const major = Number(match[1])
@@ -381,7 +402,7 @@ function supportsPromptCacheBreakpoint(model: Provider.Model): boolean {
 }
 // kilocode_change end
 
-function applyCaching(msgs: ModelMessage[], model: Provider.Model): ModelMessage[] {
+function applyCaching(msgs: ModelMessage[], model: Provider.Model, options: Record<string, unknown> = {}): ModelMessage[] { // kilocode_change
   const system = msgs.filter((msg) => msg.role === "system").slice(0, 2)
   const final = msgs.filter((msg) => msg.role !== "system").slice(-2)
 
@@ -405,7 +426,7 @@ function applyCaching(msgs: ModelMessage[], model: Provider.Model): ModelMessage
       cacheControl: { type: "ephemeral" },
     },
     // kilocode_change start
-    ...(supportsPromptCacheBreakpoint(model)
+    ...(supportsPromptCacheBreakpoint(model, options)
       ? {
           openai: {
             promptCacheBreakpoint: { mode: "explicit" },
@@ -536,11 +557,11 @@ export function message(msgs: ModelMessage[], model: Provider.Model, options: Re
       ((model.api.npm === "@ai-sdk/openai" ||
         model.api.npm === "@ai-sdk/azure" ||
         model.api.npm === "@kilocode/kilo-gateway") &&
-        supportsPromptCacheBreakpoint(model))) &&
+        supportsPromptCacheBreakpoint(model, options))) &&
     model.api.npm !== "@ai-sdk/gateway" &&
     !usesAnthropicAutomaticCaching
   ) {
-    msgs = applyCaching(msgs, model)
+    msgs = applyCaching(msgs, model, options)
   }
   // kilocode_change end
 

packages/opencode/src/session/llm.ts

@@ -320,6 +320,7 @@ const live: Layer.Layer<
           }),
           // kilocode_change end
           providerOptions: prepared.params.options,
+          messageTransformOptions: prepared.messageTransformOptions, // kilocode_change
           headers: prepared.headers,
           abort: input.abort,
         })

packages/opencode/src/session/llm/native-runtime.ts

@@ -39,6 +39,7 @@ type StreamInput = {
   readonly topK?: number
   readonly maxOutputTokens?: number
   readonly providerOptions?: Record<string, any>
+  readonly messageTransformOptions?: Record<string, any> // kilocode_change - endpoint-override-aware transform context
   readonly headers: Record<string, string>
   readonly abort: AbortSignal
 }
@@ -91,7 +92,7 @@ export function stream(input: StreamInput): StreamResult {
     model: input.model,
     apiKey: current.apiKey,
     baseURL: current.baseURL,
-    messages: ProviderTransform.message(input.messages, input.model, input.providerOptions ?? {}),
+    messages: ProviderTransform.message(input.messages, input.model, input.messageTransformOptions ?? input.providerOptions ?? {}), // kilocode_change
     toolChoice: input.toolChoice,
     temperature: input.temperature,
     topP: input.topP,

packages/opencode/src/session/llm/request.ts

@@ -229,7 +229,15 @@ export const prepare = Effect.fn("LLMRequestPrep.prepare")(function* (input: Pre
     messages,
     tools: Object.fromEntries(Object.entries(tools).toSorted(([a], [b]) => a.localeCompare(b))),
     params,
-    messageTransformOptions: options,
+    // kilocode_change start - surface provider-level endpoint overrides to message
+    // transforms without leaking them into the wire params (options is also params.options)
+    messageTransformOptions: {
+      ...options,
+      ...(typeof (input.provider.options?.endpoint ?? input.provider.options?.baseURL) === "string"
+        ? { providerEndpointOverride: input.provider.options?.endpoint ?? input.provider.options?.baseURL }
+        : {}),
+    },
+    // kilocode_change end
     headers: {
       ...(input.model.providerID.startsWith("kilo") // kilocode_change
         ? {

packages/opencode/test/kilocode/provider/transform-cache-breakpoint.test.ts

@@ -0,0 +1,132 @@
+import { describe, expect, test } from "bun:test"
+import { ProviderTransform } from "@/provider/transform"
+
+describe("ProviderTransform.message - prompt cache breakpoint endpoint gating", () => {
+  const createModel = (overrides: Partial<any> = {}) =>
+    ({
+      id: "gpt-5.6",
+      providerID: "openai",
+      api: {
+        id: "gpt-5.6",
+        url: "https://api.openai.com/v1",
+        npm: "@ai-sdk/openai",
+      },
+      name: "GPT 5.6",
+      capabilities: {
+        temperature: true,
+        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: 0.00125, output: 0.01, cache: { read: 0.000125, write: 0 } },
+      limit: { context: 400_000, output: 128_000 },
+      status: "active",
+      options: {},
+      headers: {},
+      ...overrides,
+    }) as any
+
+  const msgs = () =>
+    [
+      { role: "system", content: "You are a helpful assistant" },
+      { role: "user", content: "Hello" },
+    ] as any[]
+
+  test("custom @ai-sdk/openai provider does not apply promptCacheBreakpoint", () => {
+    const model = createModel({
+      providerID: "custom",
+      api: {
+        id: "gpt-5.6-sol",
+        url: "https://redacted/v1",
+        npm: "@ai-sdk/openai",
+      },
+      id: "gpt-5.6-sol",
+    })
+
+    const result = ProviderTransform.message(msgs(), model, {}) as any[]
+
+    expect(result[0].providerOptions?.openai?.promptCacheBreakpoint).toBeUndefined()
+    expect(result[1].providerOptions?.openai?.promptCacheBreakpoint).toBeUndefined()
+  })
+
+  test("openai provider with a custom endpoint override does not apply promptCacheBreakpoint", () => {
+    const model = createModel()
+
+    const result = ProviderTransform.message(msgs(), model, {
+      providerEndpointOverride: "https://proxy.example.com/v1",
+    }) as any[]
+
+    expect(result[0].providerOptions?.openai?.promptCacheBreakpoint).toBeUndefined()
+    expect(result[1].providerOptions?.openai?.promptCacheBreakpoint).toBeUndefined()
+  })
+
+  test("openai provider with a model-level baseURL override does not apply promptCacheBreakpoint", () => {
+    const model = createModel()
+
+    const result = ProviderTransform.message(msgs(), model, {
+      baseURL: "https://gateway.internal:8443/openai/v1",
+    }) as any[]
+
+    expect(result[0].providerOptions?.openai?.promptCacheBreakpoint).toBeUndefined()
+    expect(result[1].providerOptions?.openai?.promptCacheBreakpoint).toBeUndefined()
+  })
+
+  test("openai provider with an unparseable override does not apply promptCacheBreakpoint", () => {
+    const model = createModel()
+
+    const result = ProviderTransform.message(msgs(), model, {
+      providerEndpointOverride: "not a url",
+    }) as any[]
+
+    expect(result[0].providerOptions?.openai?.promptCacheBreakpoint).toBeUndefined()
+    expect(result[1].providerOptions?.openai?.promptCacheBreakpoint).toBeUndefined()
+  })
+
+  test("first-party openai without overrides still applies promptCacheBreakpoint", () => {
+    const model = createModel()
+
+    const result = ProviderTransform.message(msgs(), model, {}) as any[]
+
+    expect(result[0].providerOptions?.openai?.promptCacheBreakpoint).toEqual({ mode: "explicit" })
+    expect(result[1].providerOptions?.openai?.promptCacheBreakpoint).toEqual({ mode: "explicit" })
+  })
+
+  test("azure endpoint override on an azure host still applies promptCacheBreakpoint", () => {
+    const model = createModel({
+      providerID: "azure",
+      api: {
+        id: "gpt-5.6",
+        url: "",
+        npm: "@ai-sdk/azure",
+      },
+    })
+
+    const result = ProviderTransform.message(msgs(), model, {
+      providerEndpointOverride: "https://myresource.cognitiveservices.azure.com/openai/v1",
+    }) as any[]
+
+    expect(result[0].providerOptions?.azure?.promptCacheBreakpoint).toEqual({ mode: "explicit" })
+    expect(result[1].providerOptions?.azure?.promptCacheBreakpoint).toEqual({ mode: "explicit" })
+  })
+
+  test("azure endpoint override on a non-azure host does not apply promptCacheBreakpoint", () => {
+    const model = createModel({
+      providerID: "azure",
+      api: {
+        id: "gpt-5.6",
+        url: "",
+        npm: "@ai-sdk/azure",
+      },
+    })
+
+    const result = ProviderTransform.message(msgs(), model, {
+      providerEndpointOverride: "https://proxy.example.com/azure/v1",
+    }) as any[]
+
+    expect(result[0].providerOptions?.azure?.promptCacheBreakpoint).toBeUndefined()
+    expect(result[1].providerOptions?.azure?.promptCacheBreakpoint).toBeUndefined()
+  })
+})