fix(agent-core-v2): disable SDK-internal retries that blocked cancellation

.changeset/cli-retry-cancel.md

@@ -0,0 +1,5 @@
+---
+"@moonshot-ai/kimi-code": patch
+---
+
+Fix Ctrl+C being ignored during automatic retries of failed API requests (rate limits, server errors, connection failures).

packages/agent-core-v2/src/kosong/provider/bases/anthropic/anthropic.ts

@@ -20,6 +20,11 @@
  * the trait-composed `convertError` hook consulted, so a vendor riding this
  * transport classifies each RAW SDK failure exactly once before the base
  * rules run.
+ *
+ * The SDK client is built with `maxRetries: 0`: the SDK's internal backoff
+ * sleep never observes the turn's AbortSignal, so rate-limit / server /
+ * connection retry is owned by the engine's step-retry layer (observable and
+ * cancellable), never by the SDK.
  */
 
 import Anthropic, {
@@ -1148,6 +1153,7 @@ export class AnthropicChatProvider implements ChatProvider {
       authToken: null,
       baseURL: this._baseUrl ?? null,
       defaultHeaders: this._buildDefaultHeaders(apiKey),
+      maxRetries: 0,
     });
   }
 }

packages/agent-core-v2/src/kosong/provider/bases/google-genai/google-genai.ts

@@ -11,6 +11,11 @@
  * module's abort plumbing (abortPromise racing,
  * per-chunk checks, the catch guard that rethrows DOMException aborts before
  * error conversion) is self-contained by design.
+ *
+ * Error conversion recovers the server-directed retry delay from the wire
+ * body: the SDK's `ApiError` drops response headers, so the
+ * `google.rpc.RetryInfo` detail inside the stringified error body is the
+ * only carrier of that wait time.
  */
 
 import { ApiError as GoogleApiError, GoogleGenAI as GenAIClient } from '@google/genai';
@@ -626,7 +631,12 @@ const TIMEOUT_RE = /timed?\s*out|timeout|deadline/i;
 
 export function convertGoogleGenAIError(error: unknown): ChatProviderError {
   if (error instanceof GoogleApiError) {
-    return normalizeAPIStatusError(error.status, error.message);
+    return normalizeAPIStatusError(
+      error.status,
+      error.message,
+      undefined,
+      parseRetryInfoDelayMs(error.message),
+    );
   }
   if (error instanceof Error) {
     const msg = error.message;
@@ -645,6 +655,30 @@ export function convertGoogleGenAIError(error: unknown): ChatProviderError {
   return new ChatProviderError(`GoogleGenAI error: ${String(error)}`);
 }
 
+function parseRetryInfoDelayMs(message: string): number | null {
+  try {
+    const body: unknown = JSON.parse(message);
+    if (typeof body !== 'object' || body === null) return null;
+    const details = (body as { error?: { details?: unknown } }).error?.details;
+    if (!Array.isArray(details)) return null;
+    for (const detail of details) {
+      if (typeof detail !== 'object' || detail === null) continue;
+      const type = (detail as { '@type'?: unknown })['@type'];
+      if (typeof type !== 'string' || !type.endsWith('google.rpc.RetryInfo')) continue;
+      const retryDelay = (detail as { retryDelay?: unknown }).retryDelay;
+      if (typeof retryDelay !== 'string') continue;
+      const match = /^(\d+(?:\.\d+)?)s$/.exec(retryDelay.trim());
+      if (match?.[1] === undefined) continue;
+      const seconds = Number.parseFloat(match[1]);
+      if (!Number.isFinite(seconds) || seconds < 0) continue;
+      return Math.round(seconds * 1000);
+    }
+    return null;
+  } catch {
+    return null;
+  }
+}
+
 export class GoogleGenAIChatProvider implements ChatProvider {
   readonly name: string = 'google_genai';
 

packages/agent-core-v2/src/kosong/provider/bases/openai/openai-legacy.ts

@@ -22,6 +22,11 @@
  *    tool-result `extract_text` fallback and tool-declaration-only skip are
  *    handed over to the trait wholesale: every history message is
  *    base-converted, post-processed by the hook, and dropped on `null`.
+ *
+ * The SDK client is built with `maxRetries: 0`: the SDK's internal backoff
+ * sleep never observes the turn's AbortSignal, so rate-limit / server /
+ * connection retry is owned by the engine's step-retry layer (observable and
+ * cancellable), never by the SDK.
  */
 
 import OpenAI from 'openai';
@@ -751,6 +756,7 @@ export class OpenAILegacyChatProvider implements ChatProvider {
     const clientOpts: Record<string, unknown> = {
       apiKey,
       baseURL: this._baseUrl,
+      maxRetries: 0,
     };
     const defaultHeaders = mergeRequestHeaders(this._defaultHeaders, auth?.headers);
     if (defaultHeaders !== undefined) {

packages/agent-core-v2/src/kosong/provider/bases/openai/openai-responses.ts

@@ -11,6 +11,11 @@
  * classification (already-converted errors crossing an outer catch pass
  * through without re-consulting). The developer-role model detection lives
  * here.
+ *
+ * The SDK client is built with `maxRetries: 0`: the SDK's internal backoff
+ * sleep never observes the turn's AbortSignal, so rate-limit / server /
+ * connection retry is owned by the engine's step-retry layer (observable and
+ * cancellable), never by the SDK.
  */
 
 import OpenAI from 'openai';
@@ -1203,6 +1208,7 @@ export class OpenAIResponsesChatProvider implements ChatProvider {
     const clientOpts: Record<string, unknown> = {
       apiKey,
       baseURL: this._baseUrl,
+      maxRetries: 0,
     };
     const defaultHeaders = mergeRequestHeaders(this._defaultHeaders, auth?.headers);
     if (defaultHeaders !== undefined) {

packages/agent-core-v2/test/kosong/provider/composition.test.ts

@@ -33,11 +33,19 @@
  *     profile, and the OpenAI `reasoning_effort` auto-enable with its
  *     load-bearing kill switch (a `withThinking` hook disables it).
  *
+ * Plus one construction invariant: every wire base builds its SDK client
+ * with `maxRetries: 0` — retry is owned by the engine's step-retry layer,
+ * never by the SDK (whose backoff sleep ignores the turn's AbortSignal). The
+ * closing section proves it over a real HTTP 429: the first response reaches
+ * the caller after exactly one request, carrying the server-directed delay.
+ *
  * Note: base/definition registries are module-level state shared across this
  * file, so the contribs and test-vendor definitions are imported/registered
  * exactly once here.
  */
 
+import { createServer } from 'node:http';
+
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
 
 import { APIError as AnthropicAPIError } from '@anthropic-ai/sdk';
@@ -47,6 +55,7 @@ import {
   APIConnectionError,
   APIProviderQuotaExhaustedError,
   APIProviderRateLimitError,
+  APIStatusError,
   isRetryableGenerateError,
 } from '#/kosong/contract/errors';
 import type { Message } from '#/kosong/contract/message';
@@ -344,6 +353,17 @@ describe('createChatProvider', () => {
   });
 });
 
+describe('SDK-internal retry disabled (engine-owned step retry)', () => {
+  it.each([
+    { protocol: 'openai', modelName: 'gpt-4o' },
+    { protocol: 'openai_responses', modelName: 'gpt-5' },
+    { protocol: 'anthropic', modelName: 'claude-opus-4-6' },
+  ] as const)('builds the $protocol SDK client with maxRetries 0', ({ protocol, modelName }) => {
+    const provider = registry.createChatProvider({ protocol, modelName, apiKey: 'sk-probe' });
+    expect((sdkClient(provider) as { maxRetries?: number }).maxRetries).toBe(0);
+  });
+});
+
 describe('google-genai vertex mode (providerOptions)', () => {
   it('forwards vertexai + project + location from providerOptions to the base', () => {
     const provider = registry.createChatProvider({
@@ -1234,3 +1254,88 @@ describe('OpenAI reasoning_effort path (issue #1616)', () => {
     expect(explicit['reasoning_effort']).toBe('low');
   });
 });
+
+describe('429 wire behavior over real HTTP (no hidden SDK retry)', () => {
+  async function with429Server(
+    body: Record<string, unknown>,
+    run: (port: number, requestCount: () => number) => Promise<void>,
+  ): Promise<void> {
+    let count = 0;
+    const server = createServer((_req, res) => {
+      count += 1;
+      res.writeHead(429, { 'content-type': 'application/json', 'retry-after': '5' });
+      res.end(JSON.stringify(body));
+    });
+    await new Promise<void>((resolve) => {
+      server.listen(0, '127.0.0.1', resolve);
+    });
+    try {
+      const address = server.address();
+      if (address === null || typeof address === 'string') {
+        throw new Error('server has no address');
+      }
+      await run(address.port, () => count);
+    } finally {
+      await new Promise<void>((resolve) => {
+        server.close(() => {
+          resolve();
+        });
+      });
+    }
+  }
+
+  it.each([
+    {
+      protocol: 'openai',
+      modelName: 'gpt-4o',
+      baseUrlPath: '/v1',
+      body: { error: { message: 'slow down', type: 'rate_limit_error' } },
+    },
+    {
+      protocol: 'openai_responses',
+      modelName: 'gpt-5',
+      baseUrlPath: '/v1',
+      body: { error: { message: 'slow down', type: 'rate_limit_error' } },
+    },
+    {
+      protocol: 'anthropic',
+      modelName: 'claude-opus-4-6',
+      baseUrlPath: '',
+      body: { type: 'error', error: { type: 'rate_limit_error', message: 'slow down' } },
+    },
+    {
+      protocol: 'google-genai',
+      modelName: 'gemini-2.5-flash',
+      baseUrlPath: '',
+      body: {
+        error: {
+          code: 429,
+          message: 'Resource exhausted',
+          status: 'RESOURCE_EXHAUSTED',
+          details: [{ '@type': 'type.googleapis.com/google.rpc.RetryInfo', retryDelay: '5s' }],
+        },
+      },
+    },
+  ] as const)(
+    'the first 429 reaches the caller after exactly one request with the 5s server delay ($protocol)',
+    async ({ protocol, modelName, baseUrlPath, body }) => {
+      await with429Server(body, async (port, requestCount) => {
+        const provider = registry.createChatProvider({
+          protocol,
+          modelName,
+          apiKey: 'sk-probe',
+          baseUrl: `http://127.0.0.1:${String(port)}${baseUrlPath}`,
+        });
+        const rejected: unknown = await provider.generate('sys', [], PROBE_HISTORY).then(
+          () => {
+            throw new Error('expected generate to reject');
+          },
+          (error: unknown) => error,
+        );
+        expect(rejected).toBeInstanceOf(APIProviderRateLimitError);
+        expect((rejected as APIStatusError).retryAfterMs).toBe(5000);
+        expect(requestCount()).toBe(1);
+      });
+    },
+  );
+});

packages/agent-core-v2/test/kosong/provider/errors.test.ts

@@ -17,9 +17,15 @@
  * last-declarer-wins semantics and consulted — after the abort guard — with
  * the raw failure at every catch seam, including the Responses in-stream
  * error-event path.
+ *
+ * The Google GenAI converter coverage checks the recovery of the
+ * server-directed retry delay from the wire body's `google.rpc.RetryInfo`
+ * detail — the SDK's `ApiError` drops response headers, so the body detail
+ * is the only carrier of that wait time.
  */
 
 import { APIError as AnthropicAPIError } from '@anthropic-ai/sdk';
+import { ApiError as GoogleApiError } from '@google/genai';
 import { APIError as OpenAIAPIError } from 'openai';
 import { describe, expect, it } from 'vitest';
 
@@ -34,6 +40,7 @@ import {
 import type { ProtocolAdapterConfig } from '#/kosong/protocol/protocol';
 import { traitConvertError, type TraitContext } from '#/kosong/protocol/protocolTrait';
 import { convertAnthropicError } from '#/kosong/provider/bases/anthropic/anthropic';
+import { convertGoogleGenAIError } from '#/kosong/provider/bases/google-genai/google-genai';
 import { convertOpenAIError } from '#/kosong/provider/bases/openai/openai-common';
 import { OpenAIResponsesStreamedMessage } from '#/kosong/provider/bases/openai/openai-responses';
 import { composeOpenAIChatHooks } from '#/kosong/provider/bases/openai/openaiHooks';
@@ -301,3 +308,58 @@ describe('OpenAI Responses quota-exhausted conversion', () => {
     expect(seen[0]).toMatchObject({ type: 'error', code: 'vendor_quota_gone' });
   });
 });
+
+describe('convertGoogleGenAIError RetryInfo recovery', () => {
+  function googleApiError(status: number, body: unknown): GoogleApiError {
+    return new GoogleApiError({ message: JSON.stringify(body), status });
+  }
+
+  it('recovers the server-directed retry delay from the RetryInfo detail', () => {
+    const error = convertGoogleGenAIError(
+      googleApiError(429, {
+        error: {
+          code: 429,
+          message: 'Resource exhausted',
+          status: 'RESOURCE_EXHAUSTED',
+          details: [
+            { '@type': 'type.googleapis.com/google.rpc.QuotaFailure', violations: [] },
+            { '@type': 'type.googleapis.com/google.rpc.RetryInfo', retryDelay: '5s' },
+          ],
+        },
+      }),
+    );
+    expect(error).toBeInstanceOf(APIProviderRateLimitError);
+    expect((error as APIStatusError).retryAfterMs).toBe(5000);
+    expect(isRetryableGenerateError(error)).toBe(true);
+  });
+
+  it('parses fractional proto Duration delays', () => {
+    const error = convertGoogleGenAIError(
+      googleApiError(429, {
+        error: {
+          details: [
+            { '@type': 'type.googleapis.com/google.rpc.RetryInfo', retryDelay: '5.5s' },
+          ],
+        },
+      }),
+    );
+    expect((error as APIStatusError).retryAfterMs).toBe(5500);
+  });
+
+  it('keeps a 429 without RetryInfo a retryable rate limit with no server delay', () => {
+    const error = convertGoogleGenAIError(
+      googleApiError(429, { error: { code: 429, message: 'Too many requests' } }),
+    );
+    expect(error).toBeInstanceOf(APIProviderRateLimitError);
+    expect((error as APIStatusError).retryAfterMs).toBeNull();
+    expect(isRetryableGenerateError(error)).toBe(true);
+  });
+
+  it('tolerates a non-JSON ApiError message', () => {
+    const error = convertGoogleGenAIError(
+      new GoogleApiError({ message: 'Too many requests', status: 429 }),
+    );
+    expect(error).toBeInstanceOf(APIProviderRateLimitError);
+    expect((error as APIStatusError).retryAfterMs).toBeNull();
+  });
+});