fix: surface provider error details from Responses API stream failures

.changeset/responses-stream-error-details.md

@@ -0,0 +1,5 @@
+---
+"@kilocode/cli": patch
+---
+
+Show the provider's actual error message when an OpenAI or Azure Responses API stream fails (for example an upstream rate limit) instead of a generic retry notice.

package.json

@@ -54,7 +54,7 @@
       "drizzle-kit": "1.0.0-rc.2",
       "drizzle-orm": "1.0.0-rc.2",
       "effect": "4.0.0-beta.74",
-      "ai": "6.0.168",
+      "ai": "6.0.235",
       "cross-spawn": "7.0.6",
       "hono": "4.12.12",
       "hono-openapi": "1.1.2",

packages/core/package.json

@@ -85,20 +85,20 @@
     "@ai-sdk/alibaba": "1.0.17",
     "@ai-sdk/amazon-bedrock": "4.0.112",
     "@ai-sdk/anthropic": "3.0.82",
-    "@ai-sdk/azure": "3.0.49",
+    "@ai-sdk/azure": "3.0.93",
     "@ai-sdk/cerebras": "2.0.54",
     "@ai-sdk/cohere": "3.0.27",
     "@ai-sdk/deepinfra": "2.0.41",
-    "@ai-sdk/gateway": "3.0.104",
+    "@ai-sdk/gateway": "3.0.157",
     "@ai-sdk/google": "3.0.73",
     "@ai-sdk/google-vertex": "4.0.128",
     "@ai-sdk/groq": "3.0.31",
     "@ai-sdk/mistral": "3.0.27",
-    "@ai-sdk/openai": "3.0.53",
+    "@ai-sdk/openai": "3.0.88",
     "@ai-sdk/openai-compatible": "2.0.48",
     "@ai-sdk/perplexity": "3.0.26",
-    "@ai-sdk/provider": "3.0.8",
-    "@ai-sdk/provider-utils": "4.0.23",
+    "@ai-sdk/provider": "3.0.14",
+    "@ai-sdk/provider-utils": "4.0.40",
     "@ai-sdk/togetherai": "2.0.41",
     "@ai-sdk/vercel": "2.0.39",
     "@ai-sdk/xai": "3.0.102",

packages/kilo-gateway/package.json

@@ -34,8 +34,8 @@
   "dependencies": {
     "@kilocode/plugin": "workspace:*",
     "@ai-sdk/alibaba": "1.0.17",
-    "@ai-sdk/anthropic": "3.0.71",
-    "@ai-sdk/openai": "3.0.53",
+    "@ai-sdk/anthropic": "3.0.82",
+    "@ai-sdk/openai": "3.0.88",
     "@ai-sdk/openai-compatible": "2.0.48",
     "@ai-sdk/mistral": "3.0.27",
     "@openrouter/ai-sdk-provider": "2.9.0",

packages/opencode/package.json

@@ -73,20 +73,20 @@
     "@ai-sdk/alibaba": "1.0.17",
     "@ai-sdk/amazon-bedrock": "4.0.112",
     "@ai-sdk/anthropic": "3.0.82",
-    "@ai-sdk/azure": "3.0.49",
+    "@ai-sdk/azure": "3.0.93",
     "@ai-sdk/cerebras": "2.0.54",
     "@ai-sdk/cohere": "3.0.27",
     "@ai-sdk/deepinfra": "2.0.41",
-    "@ai-sdk/gateway": "3.0.104",
+    "@ai-sdk/gateway": "3.0.157",
     "@ai-sdk/google": "3.0.73",
     "@ai-sdk/google-vertex": "4.0.128",
     "@ai-sdk/groq": "3.0.31",
     "@ai-sdk/mistral": "3.0.27",
-    "@ai-sdk/openai": "3.0.53",
+    "@ai-sdk/openai": "3.0.88",
     "@ai-sdk/openai-compatible": "2.0.48",
     "@ai-sdk/perplexity": "3.0.26",
-    "@ai-sdk/provider": "3.0.8",
-    "@ai-sdk/provider-utils": "4.0.23",
+    "@ai-sdk/provider": "3.0.14",
+    "@ai-sdk/provider-utils": "4.0.40",
     "@ai-sdk/togetherai": "2.0.41",
     "@ai-sdk/vercel": "2.0.39",
     "@ai-sdk/xai": "3.0.102",

packages/opencode/src/kilocode/provider/error.ts

@@ -1,5 +1,82 @@
 import type { APICallError } from "ai"
 import { ProviderV2 } from "@opencode-ai/core/provider"
+import { isRecord } from "@/util/record"
+import { z } from "zod"
+
+export type Frame = {
+  type?: unknown
+  error?: Record<string, unknown>
+} & Record<string, unknown>
+
+const payload = z.looseObject({
+  code: z.union([z.string(), z.number()]).optional(),
+  message: z.string().optional(),
+})
+
+// OpenAI Responses API terminal frame forwarded by @ai-sdk/openai >= 3.0.82
+const terminal = z.looseObject({
+  type: z.literal("response.failed"),
+  response: z.looseObject({ error: payload }),
+})
+
+// Envelope-less chat-completions wrapper
+const wrapper = z.looseObject({
+  type: z.undefined().optional(),
+  error: payload,
+})
+
+// Bare error object; a nested `error` record wins over bare top-level
+// fields so gateway wrappers keep the specific inner code
+const bare = z.looseObject({
+  type: z.undefined().optional(),
+  code: z.union([z.string(), z.number()]),
+  message: z.string(),
+})
+
+/**
+ * Normalize provider stream error frames that arrive without the
+ * `{ type: "error" }` envelope expected by ProviderError.parseStreamError.
+ */
+export function frame(body: unknown): Frame {
+  const failed = terminal.safeParse(body)
+  if (failed.success) return { type: "error", error: failed.data.response.error }
+  const wrapped = wrapper.safeParse(body)
+  if (wrapped.success) return { ...wrapped.data, type: "error" }
+  const direct = bare.safeParse(body)
+  if (direct.success) return { ...direct.data, type: "error", error: direct.data }
+  if (!isRecord(body)) return {}
+  return { ...body, error: isRecord(body.error) ? body.error : undefined }
+}
+
+const RETRYABLE = /rate.?limit|too.?many.?requests|rate increased too quickly|exhausted|overload|server|unavailable|timeout/i
+
+// Must stay at least as permissive as the Session.retryable heuristics that
+// applied when these frames still surfaced as NamedError.Unknown, or
+// previously-retried provider errors silently become terminal
+function retryable(error: Frame["error"], message: string) {
+  const code = error?.code
+  const numeric = typeof code === "number" ? code : typeof code === "string" && code.trim() !== "" ? Number(code) : NaN
+  if (!Number.isNaN(numeric)) {
+    if (numeric === 429 || (numeric >= 500 && numeric < 600)) return true
+  } else if (typeof code === "string" && RETRYABLE.test(code)) {
+    return true
+  }
+  const type = error?.type
+  if (typeof type === "string" && RETRYABLE.test(type)) return true
+  return RETRYABLE.test(message)
+}
+
+/**
+ * Terminal handler for normalized frames whose error code is not listed in
+ * ProviderError.parseStreamError: surface the provider message instead of
+ * falling back to a raw JSON dump. Retryable only for rate-limit and
+ * 5xx-style signals in the code, type, or message.
+ */
+export function fallback(body: Frame, responseBody: string) {
+  const message = body.error?.message
+  if (typeof message !== "string" || !message.trim()) return
+  return { type: "api_error" as const, message, isRetryable: retryable(body.error, message), responseBody }
+}
 
 const AUTH_ERROR =
   "Request had invalid authentication credentials. Expected OAuth 2 access token, login cookie or other valid authentication credential. See https://developers.google.com/identity/sign-in/web/devconsole-project."

packages/opencode/src/provider/error.ts

@@ -109,10 +109,13 @@ export type ParsedStreamError =
 
 export function parseStreamError(input: unknown): ParsedStreamError | undefined {
   const raw = json(input)
-  const body = typeof raw?.message === "string" ? (json(raw.message) ?? raw) : raw
-  if (!body) return
+  // kilocode_change start - unwrap response.failed frames and bare provider error objects before the envelope gate
+  const original = typeof raw?.message === "string" ? (json(raw.message) ?? raw) : raw
+  if (!original) return
 
-  const responseBody = JSON.stringify(body)
+  const responseBody = JSON.stringify(original)
+  const body = KiloError.frame(original)
+  // kilocode_change end
   if (body.type !== "error") return
 
   switch (body?.error?.code) {
@@ -164,6 +167,7 @@ export function parseStreamError(input: unknown): ParsedStreamError | undefined
         responseBody,
       }
   }
+  return KiloError.fallback(body, responseBody) // kilocode_change - render unlisted provider error codes as clean messages
 }
 
 export type ParsedAPICallError =

packages/opencode/test/kilocode/provider/error.test.ts

@@ -78,6 +78,123 @@ describe("provider stream errors", () => {
   })
 })
 
+describe("responses api terminal frames", () => {
+  test("surfaces the provider message from a response.failed frame", () => {
+    const payload = {
+      type: "response.failed",
+      sequence_number: 8,
+      response: {
+        error: {
+          code: "rate_limit_exceeded",
+          message:
+            "openai/gpt-5.6-terra-pro is temporarily rate-limited upstream. Please retry shortly, or add your own key to accumulate your rate limits: https://openrouter.ai/settings/integrations",
+        },
+        incomplete_details: null,
+        service_tier: "auto",
+      },
+    }
+    const result = MessageV2.fromError(payload, { providerID: ProviderV2.ID.make("openai") })
+
+    expect(MessageV2.APIError.isInstance(result)).toBe(true)
+    if (!MessageV2.APIError.isInstance(result)) throw new Error("expected APIError")
+    expect(result.data.message).toBe(payload.response.error.message)
+    expect(result.data.isRetryable).toBe(true)
+    expect(result.data.responseBody).toBe(JSON.stringify(payload))
+  })
+
+  test("unwraps bare chat-completions error objects", () => {
+    const payload = { code: "rate_limit_exceeded", message: "Try again in 30 seconds." }
+    const result = MessageV2.fromError(payload, { providerID: ProviderV2.ID.make("openai") })
+
+    expect(MessageV2.APIError.isInstance(result)).toBe(true)
+    if (!MessageV2.APIError.isInstance(result)) throw new Error("expected APIError")
+    expect(result.data.message).toBe(payload.message)
+    expect(result.data.isRetryable).toBe(true)
+  })
+
+  test("unwraps envelope-less error wrappers", () => {
+    const payload = { error: { code: "rate_limit_exceeded", message: "Try again in 30 seconds.", type: "tokens" } }
+    const result = MessageV2.fromError(payload, { providerID: ProviderV2.ID.make("openai") })
+
+    expect(MessageV2.APIError.isInstance(result)).toBe(true)
+    if (!MessageV2.APIError.isInstance(result)) throw new Error("expected APIError")
+    expect(result.data.message).toBe(payload.error.message)
+    expect(result.data.isRetryable).toBe(true)
+  })
+
+  test("prefers a nested error record over bare top-level fields", () => {
+    const payload = {
+      code: 429,
+      message: "gateway wrapper",
+      error: { code: "rate_limit_exceeded", message: "inner provider detail" },
+    }
+    const result = MessageV2.fromError(payload, { providerID: ProviderV2.ID.make("openai") })
+
+    expect(MessageV2.APIError.isInstance(result)).toBe(true)
+    if (!MessageV2.APIError.isInstance(result)) throw new Error("expected APIError")
+    expect(result.data.message).toBe(payload.error.message)
+    expect(result.data.isRetryable).toBe(true)
+  })
+
+  test("surfaces messages for unlisted error codes", () => {
+    const payload = {
+      type: "response.failed",
+      response: { error: { code: "model_not_found", message: "no such model" }, incomplete_details: null },
+    }
+    const result = MessageV2.fromError(payload, { providerID: ProviderV2.ID.make("openai") })
+
+    expect(MessageV2.APIError.isInstance(result)).toBe(true)
+    if (!MessageV2.APIError.isInstance(result)) throw new Error("expected APIError")
+    expect(result.data.message).toBe("no such model")
+    expect(result.data.isRetryable).toBe(false)
+  })
+
+  test("marks numeric 429 and 5xx error codes retryable", () => {
+    for (const payload of [
+      { code: 429, message: "too many requests" },
+      { code: 503, message: "service unavailable" },
+    ]) {
+      const result = MessageV2.fromError(payload, { providerID: ProviderV2.ID.make("openai") })
+
+      expect(MessageV2.APIError.isInstance(result)).toBe(true)
+      if (!MessageV2.APIError.isInstance(result)) throw new Error("expected APIError")
+      expect(result.data.message).toBe(payload.message)
+      expect(result.data.isRetryable).toBe(true)
+    }
+  })
+
+  test("retries exhausted and anthropic rate-limit shapes like the old heuristics", () => {
+    const exhausted = { code: "resource_exhausted", message: "quota exceeded for the day" }
+    const anthropic = {
+      type: "error",
+      error: { type: "rate_limit_error", message: "request has exceeded your per-minute rate limit" },
+    }
+    for (const payload of [exhausted, anthropic]) {
+      const result = MessageV2.fromError(payload, { providerID: ProviderV2.ID.make("openai") })
+
+      expect(MessageV2.APIError.isInstance(result)).toBe(true)
+      if (!MessageV2.APIError.isInstance(result)) throw new Error("expected APIError")
+      expect(result.data.isRetryable).toBe(true)
+    }
+  })
+
+  test("marks numeric-string error codes retryable", () => {
+    const payload = { code: "429", message: "slow down" }
+    const result = MessageV2.fromError(payload, { providerID: ProviderV2.ID.make("openai") })
+
+    expect(MessageV2.APIError.isInstance(result)).toBe(true)
+    if (!MessageV2.APIError.isInstance(result)) throw new Error("expected APIError")
+    expect(result.data.isRetryable).toBe(true)
+  })
+
+  test("ignores response.failed frames without an error payload", () => {
+    const payload = { type: "response.failed", response: { error: null, incomplete_details: null } }
+    const result = MessageV2.fromError(payload, { providerID: ProviderV2.ID.make("openai") })
+
+    expect(MessageV2.APIError.isInstance(result)).toBe(false)
+  })
+})
+
 describe("Google Gemini authentication errors", () => {
   test("explains how to troubleshoot the rejected API key", () => {
     const error = apiError(googleAuthError, "ACCESS_TOKEN_TYPE_UNSUPPORTED")